Global Aspects of Ergodic Group Actions Alexander S. Kechris
To Olympia, Sotiri, Katerina, David and Alexi
Contents ...
14 downloads
697 Views
1MB Size
Report
This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
Report copyright / DMCA form
Global Aspects of Ergodic Group Actions Alexander S. Kechris
To Olympia, Sotiri, Katerina, David and Alexi
Contents Introduction
iii
Chapter I. Measure preserving automorphisms 1. The group Aut(X, µ) 2. Some basic facts about Aut(X, µ) 3. Full groups of equivalence relations 4. The Reconstruction Theorem 5. Turbulence of conjugacy 6. Automorphism groups of equivalence relations 7. The outer automorphism group 8. Costs and the outer automorphism group 9. Inner amenability
1 1 5 13 20 30 40 44 48 52
Chapter II. The space of actions 10. Basic properties 11. Characterizations of groups with property (T) and HAP 12. The structure of the set of ergodic actions 13. Turbulence of conjugacy in the ergodic actions 14. Conjugacy in ergodic actions of property (T) groups 15. Connectedness in the space of actions 16. The action of SL2 (Z) on T2 17. Non-orbit equivalent actions of free groups 18. Classifying group actions: A survey
61 61 79 84 90 98 105 113 117 122
Chapter III. Cocycles and cohomology 19. Group-valued random variables 20. Cocycles 21. The Mackey action and reduction cocycles 22. Homogeneous spaces and Effros’ Theorem 23. Isometric actions 24. Topology on the space of cocycles 25. Cohomology I: Some general facts 26. Cohomology II: The hyperfinite case 27. Cohomology III: The non E0 -ergodic case 28. The minimal condition on centralizers 29. Cohomology IV: The E0 -ergodic case 30. Cohomology V: Actions of property (T) groups
125 125 129 135 143 145 148 150 153 158 168 170 179
i
ii
CONTENTS
Appendix A.
Realifications and complexifications
187
Appendix B.
Tensor products of Hilbert spaces
189
Appendix C. Gaussian probability spaces
191
Appendix D.
Wiener chaos decomposition
193
Appendix E.
Extending representations to actions
197
Appendix F.
Unitary representations of abelian groups
199
Appendix G.
Induced representations and actions
203
Appendix H.
The space of unitary representations
207
Appendix I. Semidirect products of groups
223
Bibliography
227
Index
233
Introduction (A) The original motivation for this work came from the study of certain results in ergodic theory, primarily, but not exclusively, obtained over the last several years. These included: i) Work of Hjorth and ForemanWeiss concerning the complexity of the problem of classification of ergodic measure preserving transformations up to conjugacy, ii) Various results concerning the structure of the outer automorphism group of a countable measure preserving equivalence relation, including, e.g., work of Jones-Schmidt, iii) Ergodic theoretic characterizations of groups with property (T) or the Haagerup Approximation Property, in particular results of Schmidt, ConnesWeiss, Jolissaint, and Glasner-Weiss, iv) Results of Hjorth, Popa, GaboriauPopa and T¨ornquist on the existence of many non-orbit equivalent ergodic actions of certain non-amenable groups, v) Popa’s recent work on cocycle superrigidity. Despite the apparent diversity of the subjects treated in these works, we gradually realized that they can be understood within a rather unified framework. This is the study of the global structure of the space A(Γ, X, µ) of measure preserving actions of a countable group Γ on a standard measure space (X, µ) and the canonical action of the automorphism group Aut(X, µ) of (X, µ) by conjugation on A(Γ, X, µ) as well as the study of the global structure of the space of cocycles and certain canonical actions on it. Our goal here is to explore this point of view by presenting (a) earlier results, sometimes in new formulations or with new proofs, (b) new theorems, and finally (c) interesting open questions that are suggested by this approach. (B) The book is divided into three chapters, the first consisting of Sections 1–9, the second of Sections 10–18, and the third of Sections 19–30. There are also nine appendices. In the first chapter, we study the automorphism group Aut(X, µ) of a standard measure space (i.e., the group of measure preserving automorphisms of (X, µ)) and various subgroups associated with measure preserving equivalence relations. Note that Aut(X, µ) can be also identified with the space A(Z, X, µ) of measure preserving Z-actions. Sections 1, 2 review some basic facts about the group Aut(X, µ). In Section 2 we also show that the class of mild mixing transformations in Aut(X, µ) is co-analytic but not Borel (a result also proved independently by Robert Kaufman). This is in contrast with the well-known fact that the ergodic, weak mixing and (strong) mixing transformations are, resp., Gδ , Gδ and Fσδ sets. In Section iii
iv
INTRODUCTION
3 we discuss the full group [E] ⊆ Aut(X, µ) of a measure preserving countable Borel equivalence relation E on (X, µ). In Section 4 we give a detailed proof of Dye’s reconstruction theorem, which asserts that the equivalence relation E is determined up to (measure preserving) isomorphism by [E] as an abstract group. In Section 5 we give a new method for proving the turbulence property of the conjugacy action of Aut(X, µ) on the set ERG of ergodic transformations in Aut(X, µ), originally established by ForemanWeiss, and use this method to prove other turbulence results in the context of full groups. We also extend the work of Hjorth and Foreman-Weiss on non-classification by countable structures of weak mixing transformations in Aut(X, µ), up to conjugacy or unitary (spectral) equivalence, to the case of mixing transformations (and obtain more precise information in this case). In Sections 6, 7 we review the basic properties of the automorphism group N [E] of a measure preserving countable Borel equivalence relation E on (X, µ) and its outer automorphism group, Out(E), and establish (in Section 7) the turbulence of the latter, when E is hyperfinite (and even in more general situations). Understanding when Out(E) is a Polish group (with respect to the canonical topology discussed in Section 7) is an interesting open problem raised in work of Jones-Schmidt. In Section 8 we establish a connection between the Polishness of Out(E) and Gaboriau’s theory of costs and use it to obtain a partial answer to this problem. In Section 9 we discuss Effros’ notion of inner amenability and its relationship with the open problem of Schmidt of whether this property of a group is characterized by the failure of Polishness of the outer automorphism group of some equivalence relation induced by a free, measure preserving, ergodic action of the group. Some known and new results related to this concept and Schmidt’s problem are presented here. In the second chapter, we start (in Section 10) with establishing some basic properties of the space A(Γ, X, µ) of measure preserving actions of a countable group Γ on (X, µ). For finitely generated groups, we also calculate an upper bound for the descriptive complexity of the cost function on this space and use this to show that the generic action realizes the cost of the group. In Section 11 we recall several known ergodic theoretic characterizations of groups with property (T) or the Haagerup Approximation Property (HAP). In Section 12 we study the structure of the space of ergodic actions (and some of its subspaces) in A(Γ, X, µ), recasting in this context characterizations of Glasner-Weiss and Glasner concerning property (T) or the HAP, originally formulated in terms of the structure of extreme points in the space of invariant measures. In Section 13 we study, using the method introduced in Section 5, turbulence of conjugacy in the space of actions and also discuss work of Hjorth and Foreman-Weiss concerning non-classification by countable structures of such actions, up to conjugacy. We also prove an analogous result for unitary (spectral) equivalence. In Section 14, we present the essence of Hjorth’s result on the existence of many
INTRODUCTION
v
non-orbit equivalent actions of property (T) groups as a basic property concerning the topological structure of the conjugacy classes of ergodic actions of such groups. We use Hjorth’s method to show that for such groups the set of ergodic actions is clopen in the uniform topology and so is each conjugacy class of ergodic actions. In Section 15 we study connectedness properties in the space of actions, using again the method of Section 5. This illustrates the close connection between local connectedness properties and turbulence. We show, in particular, that the space A(Γ, X, µ) is path-connected in the weak topology. We also contrast this to the work in Section 14 to point out the interesting phenomenon that connectedness properties in the space of actions of a group seem to be related to properties of the group, such as amenability or property (T). In particular, for groups Γ with property (T), we determine completely the path components of the space A(Γ, X, µ) in the uniform topology. In Section 16 we discuss results of Popa concerning the action of SL2 (Z) on T2 . These are used in Section 17, along with other ideas, in the proof of a non-classification result of T¨ornquist for orbit equivalence of actions of non-abelian free groups. We also briefly discuss very recent results of Ioana, Epstein and Epstein-Ioana-Kechris-Tsankov that extend this to arbitrary non-amenable groups. Finally, Section 18 contains a survey of classification problems concerning group actions. In the third chapter, we give in Section 19 a short introduction to the properties of the group of group-valued random variables and then in Sections 20, 21 we discuss the space of cocycles of a group action or an equivalence relation and some of the invariants associated with such cocycles, like the associated Mackey action and the essential range. We also discuss cocycles arising from reductions and homomorphisms of equivalence relations. Our primary interest is in cocycles with countable (discrete) targets. The next two Sections 22 and 23 contain background material concerning continuous and isometric group actions and Effros’ Theorem. The topology of the space of cocycles is discussed in Section 24 and the study of the global properties of the cohomology equivalence relation is the subject of the final Sections 25–30. Section 25 contains some general properties of the cohomology relation, and Section 26 is concerned with the hyperfinite case. There is a large literature here but it is not our main focus in this work. There is a fundamental dichotomy in the structure of the cohomology relation for the cocycles of a given equivalence relation (or action), which in some form is already present in the work of Schmidt for the case of cocycles with abelian targets. In a precise sense, that is explained in these sections, when an equivalence relation E is non E0 -ergodic (or not strongly ergodic), the structure of the cohomology relation on its cocycles is very complicated. This is the subject in Section 27. On the other hand, when the equivalence relation E is E0 -ergodic (or strongly ergodic), then the cohomology relation is simple, i.e., smooth, provided the target groups satisfy the so-called minimal condition on centralizers, discussed in Section 28 (and these include, e.g., the abelian and the linear groups). If the target groups do not satisfy
vi
INTRODUCTION
this condition, in certain situations, like, e.g., when E is given by an action of the free group with infinitely many generators, the cohomology relation is not smooth. Thus in the E0 -ergodic case, there is an additional dichotomy having to do with the structure of the target groups. This is the topic of Section 29. Finally in Section 30 we deal with the special case of actions of groups with property (T) and discuss some recent results of Popa on cocycle superrigidity. We also establish in the above sections some new characterizations of amenable and property (T) groups, that, in particular, extend earlier results of Schmidt and also show that there is a positive link between the E0 -ergodicity of an equivalence relation E and the Polishness of its outer automorphism group, Out(E), an issue raised by Jones-Schmidt. They have pointed out that E0 -ergodicity does not in general imply that Out(E) is Polish but we show in Section 29 that one has a positive implication when E is induced by a free action of a group with the minimal condition on centralizers. Appendices A – I present background material concerning Hilbert spaces and tensor products, Gaussian probability spaces and the Wiener chaos decomposition, several relevant aspects of the theory of unitary representations (including unitary representations of abelian groups and induced representations as well as some basic results about the space of unitary representations of a group) and finally semidirect products of groups. We also include, in Appendix E, a detailed proof of the standard result that any unitary representation of a countable group is a subrepresentation of the Koopman representation associated with some measure preserving action of that group. (C) I would like to thank Mikl´os Ab´ert, Scot Adams, Oleg Ageev, Sergey Bezuglyi, Andr´es Caicedo, Clinton Conley, Damien Gaboriau, Sergey Gefter, Eli Glasner, Greg Hjorth, Adrian Ioana, John Kittrell, Alain Louveau, Ben Miller, Sorin Popa, Dinakar Ramakrishnan, Sergey Sinelshchikov, Simon Thomas, Todor Tsankov and Vladimir Zhuravlev for their comments on this book or useful conversations on matters related to it. Finally, I would like to thank Leona Kershaw for typing the manuscript. Research for this work was partially supported by NSF Grant DMS0455285.
CHAPTER I
Measure preserving automorphisms 1. The group Aut(X, µ) (A) By a standard measure space (X, µ) we mean a standard Borel space X with a non-atomic probability Borel measure µ. All such spaces are isomorphic to ([0, 1], λ), where λ is Lebesgue measure on the Borel subsets of [0,1]. Denote by MALGµ the measure algebra of µ, i.e., the algebra of Borel subsets of X, modulo null sets. It is a Polish Boolean algebra under the topology given by the complete metric d(A, B) = dµ (A, B) = µ(A∆B), where ∆ is symmetric difference. Throughout we write L2 (X, µ) = L2 (X, µ, C) for the Hilbert space of complex-valued square-integrable functions. When we want to refer explicitly to the space of real-valued functions, we will write it as L2 (X, µ, R). We denote by Aut(X, µ) the group of Borel automorphisms of X which preserve the measure µ and in which we identify two such automorphisms if they agree µ-a.e. Convention. In the sequel, we will usually ignore null sets, unless there is a danger of confusion. There are two fundamental group topologies on Aut(X, µ): the weak and the uniform topology, which we now proceed to describe. (B) The weak topology on Aut(X, µ) is generated by the functions T 7→ T (A), A ∈ MALGµ , (i.e., it is the smallest topology in which these maps are continuous). With this topology, denoted by w, (Aut(X, µ), w) is a Polish topological group. A left-invariant compatible metric is given by X δw (S, T ) = 2−n µ(S(An )∆T (An )), where {An } is a dense set in MALGµ (e.g., an algebra generating the Borel sets of X), and a complete compatible metric by δ¯w (S, T ) = δw (S, T ) + δw (S −1 , T −1 ). We can also view Aut(X, µ) as the automorphism group of the measure algebra (MALGµ , µ) (equipped with the pointwise convergence topology), as well as the closed subgroup of the isometry group Iso(MALGµ , dµ ) (again 1
2
I. MEASURE PRESERVING AUTOMORPHISMS
equipped with the pointwise convergence topology) consisting of all T ∈ Iso(MALGµ , dµ ) with T (∅) = ∅. (We only need to verify that an isometry T with T (∅) = ∅ is an automorphism of the Boolean algebra MALGµ . First notice that µ(A) = dµ (∅, A) = dµ (∅, T (A)) = µ(T (A)). Next we see that if A, B ∈ MALGµ and A ∩ B = ∅, then T (A) ∩ T (B) = ∅. Indeed dµ (A, B) = µ(A) + µ(B), thus dµ (T (A), T (B)) = dµ (A, B) = µ(A) + µ(B) = µ(T (A)) + µ(T (B)), so T (A) ∩ T (B) = ∅. From this we obtain that T (∼ A) = ∼ T (A) (where ∼ A = X \ A), since dµ (A, ∼ A) = 1, so dµ (T (A), T (∼ A)) = 1 and T (A) ∩ T (∼ A)) = ∅, so T (∼ A) =∼ T (A). Finally, A ⊆ B ⇔ A ∩ (∼ B) = ∅ ⇔ T (A) ∩ (∼ T (B)) = ∅ ⇔ T (A) ⊆ T (B), so T is an automorphism of the Boolean algebra.) Identifying T ∈ Aut(X, µ) with the unitary operator on L2 (X, µ) defined by UT (f ) = f ◦ T −1 , we can identify Aut(X, µ) (as a topological group) with a closed subgroup of U (L2 (X, µ)), the unitary group of the Hilbert space L2 (X, µ), equipped with the weak topology, which coincides on the unitary group with the strong topology. With this identification, Aut(X, µ) becomes the group of U ∈ U (L2 (X, µ)) that satisfy: U (f g) = U (f )U (g) (pointwise multiplication) whenever f, g, f g ∈ L2 (X, µ), i.e., the so-called multiplicative operators. It also coincides with the group of U ∈ U (L2 (X, µ)) that satisfy f ≥ 0 ⇒ U (f ) ≥ 0, U (1) = 1, i.e., the so-called positivity preserving operators fixing 1. Note also that any such U preserves real functions. R Let L20 (X, µ) = C⊥ = {f ∈ L2 (X, µ) : f dµ = 0} be the orthogonal of the space C = C1 of the constant functions. Clearly UT is the identity on C and L20 (X, µ) is invariant under UT . So we can also identify T with UT0 = UT |L20 (X, µ) and view Aut(X, µ) as a closed subgroup of U (L20 (X, µ)). The map T 7→ UT of Aut(X, µ) into U (L2 (X, µ)) is called the Koopman representation of Aut(X, µ). It has no non-trivial closed invariant subspaces in L20 (X, µ), i.e., the Koopman representation on L20 (X, µ) is irreducible (see Glasner, [Gl2], 5.14). Finally, for further reference, let us describe explicitly some open bases for (Aut(X, µ), w). The sets of the form VS,A1 ...An , = {T : ∀i ≤ n(µ(T (Ai )∆S(Ai )) < )}, where A1 , . . . , An ∈ MALGµ , > 0, S ∈ Aut(X, µ), form an open basis for w. This is immediate from the definition of this topology. Next we claim that the sets of the form WS,A1 ...An , = {T : ∀i, j ≤ n(|µ(S(Ai ) ∩ Aj ) − µ(T (Ai ) ∩ Aj )| < )},
1. THE GROUP Aut(X, µ)
3
for A1 , . . . , An ∈ MALGµ , > 0, S ∈ Aut(X, µ), form an open basis for w. The easiest way to see this, as pointed out by O. Ageev, is to use the identification of (Aut(X, µ), w) with a closed subgroup of U (L2 (X, µ)) with is generated by the functions U 7→ hU (f ), gi = Rthe weak topology, which U (f )gdµ, for f, g ∈ L2 (X, µ). Simply notice that if we denotePby χA the characteristic function of a set A, then the linear combinations ni=1 αi χAi of characteristic functions of Borel a dense set in the Hilbert space R sets form 2 −1 L (X, µ) and hUT (χA ), χB i = (χA ◦ T )χB dµ = µ(T (A) ∩ B). It is clear that we can restrict in VS,A1 ...An, or WS,A1 ...An, the sets A1 , . . . , An to any countable dense set in MALGµ and to rationals and still obtain bases. Also it is easy to check that we can restrict above the A1 , . . . , An to belong to any countable dense (Boolean) subalgebra of MALGµ and moreover assume that A1 , . . . , An form a partition of X (by looking, for each A1 , . . . , An , at the atoms of the Boolean algebra generated by A1 , . . . , An ). (C) We now come to the uniform topology, u. This is defined by the metric δu0 (S, T ) = sup µ(S(A)∆T (A)). A∈MALGµ
This is a 2-sided invariant complete metric on Aut(X, µ) but it is not separable (e.g., if Sα denotes the rotation by α ∈ T of the group T, then {Sα } is discrete). Clearly w $ u. An equivalent to δu0 metric is defined by δu (S, T ) = µ({x : S(x) 6= T (x)}). In fact, 2 δu ≤ δu0 ≤ δu . 3 The metric δu is also 2-sided invariant. For T : X → X let supp(T ) = {x : T (x) 6= x}. Then δu (T, 1) = µ(supp(T )), where 1 = id (the identity function on X). We note that each closed δu0 -ball is closed in w and therefore each open 0 δu -ball is Fσ in w. To see this, note that for each T0 ∈ Aut(X, µ), > 0, the ball {T : δu0 (T0 , T ) ≤ } coincides with the set of T satisfying ∀A ∈ MALGµ (dµ (T (A), T0 (A)) ≤ ). Similarly, each closed δu -ball is closed in w. To see this, it is enough to show, for each > 0, that the set {T : δu (T, 1) > } is open in w. Fix T with δP A1 , . . . , An with u (T, 1) > . Then we can find pairwise disjoint Borel setsP n n 1 µ(A ) > and T (A ) ∩ A = ∅, ∀i ≤ n. Fix δ < ( i i i i=1 i=1 µ(Ai ) − ). n Then if µ(S(Ai )∆T (Ai )) < δ, ∀i ≤ n, we have µ(S(Ai )∆Ai ) ≥ µ(T (Ai )∆Ai ) − µ(S(Ai )∆T (Ai )) > 2µ(Ai ) − δ,
4
I. MEASURE PRESERVING AUTOMORPHISMS
Pn so Pnµ(Ai \ S(Ai )) > µ(Ai ) − δ, ∀i ≤ n. Then δu (S, 1) ≥ i=1 µ(Ai \ S(Ai )) > i=1 µ(Ai ) − nδ > . It follows that both δu and δu0 are lower-semicontinuous in the space (Aut(X, µ), w)2 . Let us finally observe that similar calculations show that the map T 7→ supp(T ) Aut(X, µ) → MALGµ is continuous in (Aut(X, µ), u) and Baire class 1 in (Aut(X, µ), w). This is clear for u. For the weak topology, we need to show that for each A ∈ MALGµ , > 0, the set {T : µ(supp(T )∆A) < } = F is Fσ . Note that T ∈ F ⇔ ∃α, β[α, β ∈ Q and α, β ≥ 0 and α + β < and µ({x 6∈ A : T (x) 6= x}) ≤ α and µ({x ∈ A : T (x) = x}) < β]. Now the previous argument also shows that for each γ > 0, B ∈ MALGµ , {T : µ({x ∈ B : T (x) 6= x}) > γ} is open in w. It follows that {T : µ({x 6∈ A : T (x) 6= x}) ≤ α} is closed in w and since µ({x ∈ A : T (x) = x}) = µ(A) − µ({x ∈ A : T (x) 6= x}), the set {T : µ({x ∈ A : T (x) = x}) < β} is open in w, thus F is Fσ in w.
2
Remark. One could also consider the operator norm topology on the group Aut(X, µ) induced by the identification T 7→ UT . However, it is easy to see that this is the discrete topology. We will show that T 6= 1 ⇒ kUT − Ik ≥ 1 (where I is the identity operator on L2 (X, µ)). Since T 6= 1 there is a set of positive measure on which T (x) 6= x, so there is a set A of positive measure, say , such that T (A) ∩ A = ∅. Let f ∈ L2 (X, µ) be equal to √1 on A and 0 otherwise. Thus kf k = 1. Also kUT (f ) − f k2 = R R |f (T −1 (x)) − f (x)|2 dµ(x) ≥ x∈A |f (T −1 (x)) − f (x)|2 dµ(x) = 1. Convention. When we consider Aut(X, µ) as a topological group without explicitly indicating which topology we use, it will be assumed that it is equipped with the weak topology. Comments. The basic facts about the topologies w, u can be found in Halmos [Ha]. Note however that Halmos uses δu0 for our δu (and vice versa). Since we more often use (our) δu we decided to leave it unprimed. For the characterizations of the unitary operators UT , T ∈ Aut(X, µ), see, e.g., Fleming-Jamison [FJ], p. 73, Glasner [Gl2], p. 366, and Walters [Wa], Ch. 2.
2. SOME BASIC FACTS ABOUT Aut(X, µ)
5
2. Some basic facts about Aut(X, µ) (A) An element T ∈ Aut(X, µ) is periodic if all its orbits are finite a.e. It is periodic of period n if for almost all x, T 0 (x) = x, T 1 (x), . . . , T n−1 (x) are distinct but T n (x) = x. It is aperiodic if T n (x) 6= x for all n 6= 0, a.e. Consider now X = 2N , µ = the usual product measure. A basic nbhd of rank n is a set of the form Ns = {x ∈ 2N : s ⊆ x}, where s ∈ 2n . A dyadic permutation of rank n is determined by a permutation π of 2n via Tπ (sˆx) = π(s)ˆx. It is clearly in Aut(X, µ). If π is cyclic, we call the corresponding Tπ a cyclic dyadic permutation of rank n. In that case Tπ is periodic of period 2n . We now have the following approximation theorem. Theorem 2.1 (Weak Approximation Theorem, Halmos [Ha]). The cyclic dyadic permutations are dense in (Aut(X, µ), w). In fact, every open set contains cyclic dyadic permutations of any sufficiently high rank. Remark. The following nice argument for the special case of 2.1 that states that a measure preserving homeomorphism of X can be weakly approximated by a dyadic permutation comes from [LS], where it is attributed to Steger. Fix T ∈ Aut(X, µ), T a homeomorphism. It is enough to show that for any given > 0 and f1 , . . . , fm ∈ C(X), there is n such that for any N ≥ n, there is π of rank N such that |f` (Tπ−1 (x)) − f` (T −1 (x))| < , ∀x ∈ X, ` = 1, . . . , m. Choose n large enough so that the oscillation of every f` , f` ◦ T −1 on each basic nbhd of rank n is < /2. Fix now N ≥ n and enumerate the basic nbhds of rank N as D1 , . . . , D2N . Next define a bipartite graph G on two disjoint copies of {1, . . . , 2N } by: (i, j) ∈ G ⇔ Di ∩ T (Dj ) 6= ∅. Claim. This satisfies the criterion for the application of the marriage theorem. Granting this, by the marriage theorem, there is a bijection ϕ : {1, . . . , 2N } → {1, . . . , 2N } with Di ∩ T (Dϕ(i) ) 6= ∅. Consider now x ∈ X. Say x ∈ Di , so that Tϕ (x) ∈ Dϕ(i) . Also T −1 (Di ) ∩ Dϕ(i) 6= ∅, so find z ∈ Di with T −1 (z) = y ∈ Dϕ(i) Then |f` (Tϕ (x)) − f` (y)| < /2 and |f` (y) − f` (T −1 (x))| < /2, thus |f` (Tϕ (x)) − f` (T −1 (x))| < . Finally, put π = ϕ−1 . Proof of the claim. Fix A ⊆ {1, . . . , 2N } of cardinality M . We will check that there are ≥ M points adjacent to elements of A. If not, then there is B ⊆ {1, . . . , 2N } of cardinality >S 2N − M such that Di ∩ TP(Dj ) = ∅ for S all i ∈ A, j ∈ B. Then ( i∈A Di ) ∩ ( j∈B T (Dj )) = ∅, so 1 ≥ i∈A µ(Di ) +
6
I. MEASURE PRESERVING AUTOMORPHISMS
P µ(T (Dj )) = 2MN + j∈B µ(Dj ) > Similarly reversing the roles of i, j. P
j∈B
M 2N
+
2N −M 2N
= 1, a contradiction. 2
In particular, 2.1 implies that (Aut(X, µ), w) is topologically locally finite, i.e., has a locally finite countable dense subgroup. We also have the following result which is a consequence of the Rokhlin Lemma. Theorem 2.2 (Uniform Approximation Theorem, Rokhlin, Halmos [Ha]). If T ∈ Aut(X, µ) is aperiodic, then for each N ≥ 1, > 0 there is a periodic S ∈ Aut(X, µ) of period N such that δu (S, T ) ≤ N1 + . Consider now the set APER of all aperiodic elements of Aut(X, µ). Then as
1 T ∈ APER ⇔ ∀n (δu (T , 1) = 1) ⇔ ∀n∀m δu (T , 1) > 1 − m n
n
,
APER is Gδ in (Aut(X, µ), w). (It is also clearly closed in (Aut(X, µ), u).) In fact the following holds. Proposition 2.3. APER is dense Gδ in (Aut(X, µ), w). Proof. Take X = 2N with the usual measure µ. Then the sets of the form k \ {T : dµ (T (Di ), T0 (Di )) < }, i=1
Di finite unions of basic nbhds, T0 ∈ Aut(X, µ), > 0 form a nbhd basis for T0 in w. Thus it is enough for each dyadic permutation π, of rank say n, to find T ∈ APER such that for each s ∈ 2n , T (Ns ) = Tπ (Ns ) = Nπ(s) . Let ϕ : 2N → 2N be any aperiodic element of Aut(X, µ), e.g., the odometer: ϕ(1nˆ0ˆx) = 0nˆ1ˆx, ϕ(1∞ ) = 0∞ . Define then T (sˆx) = π(s)ˆϕ(x). Clearly this works. 2 Moreover, using the Uniform Approximation Theorem, one obtains the following result. Theorem 2.4 (Conjugacy Lemma, Halmos [Ha]). Let T ∈ APER. Then its conjugacy class {ST S −1 : S ∈ Aut(X, µ)} is uniformly dense in APER. Therefore its conjugacy class is weakly dense in Aut(X, µ). Conversely, if the conjugacy class of T is weakly dense, T ∈ APER. Proof. Observe that if U, V are periodic of period N , then they are conjugate. Indeed let A ⊆ X be a Borel set that meets (almost) every orbit of U in exactly one point and let B ⊆ X be defined similarly for V . Then µ(A) = µ(B) = N1 . Thus there is S0 ∈ Aut(X, µ) with S0 (A) = B. Define now S ∈ Aut(X, µ) by S(x) = S0 (x), if x ∈ A, n
S(U (x)) = V n (S0 (x)), if x ∈ A, 0 ≤ n < N.
2. SOME BASIC FACTS ABOUT Aut(X, µ)
7
Then S ∈ Aut(X, µ) and SU S −1 = V . Now given aperiodic T, W and > 0, find periodic U, V of large enough period N such that δu (T, U ) < , δu (W, V ) < and find S ∈ Aut(X, µ) with SU S −1 = V . Then as δu is 2-sided invariant, δu (ST S −1 , W ) ≤ δu (ST S −1 , SU S −1 ) + δu (V, W ) < 2, so we are done. For the last assertion, note that if T is not aperiodic, then for some n 6= 0, < 1, δu (T n , 1) ≤ and so δu ((ST S −1 )n , 1) ≤ for every S ∈ Aut(X, µ). 2 On the other hand every conjugacy class is small in the sense of category. In fact one has a stronger conclusion concerning unitary (or spectral) equivalence classes. Consider Aut(X, µ) as a closed subgroup of U (L2 (X, µ)) via the identification T 7→ UT . Then we say that T, S ∈ Aut(X, µ) are unitarily equivalent or spectrally equivalent if they are conjugate in U (L2 (X, µ)). This is clearly a coarser equivalence relation than conjugacy in Aut(X, µ). One now has: Theorem 2.5 (Rokhlin). Every unitary equivalence class in Aut(X, µ) is meager in the weak topology. In particular, this is true for every conjugacy class in Aut(X, µ). Proof (Hjorth). We work below in the weak topology. Fix a set A ⊆ X with µ(A) = 21 . We claim that the open set {(S, T ) : ∃n[µ(S n (A)∆A) <
1 1 and µ(T n (A)∆A) > ]} 400 9
which is equal to 1 1 and kUTn (χA ) − χA k > ]} 20 3 2 is dense in Aut(X, µ) . To see this, notice that by 2.4 the set of mixing T ∈ Aut(X, µ) is dense in Aut(X, µ). So, given any (S0 , T0 ) ∈ Aut(X, µ)2 , we can find mixing T as close as we want to T0 . Then µ(T n (A) \ A) → µ(A)µ(X \ A) = 41 , so for some N and all n ≥ N, µ(T n (A)∆A) > 91 . On the other hand, by 2.1 we can find periodic S of period ≥ N as close as we want to S0 . For such S, if S has period n, µ(S n (A)∆A) = 0, and we are done. (Notice that it would be enough here to take T weak mixing, since then there would be a set A ⊆ N of density 1 such that µ(T n (A)∆A)) > 91 , ∀n ∈ A (see, e.g., Bergelson-Gorodnik [BGo], 1.1).) Now, by 2.4, if a unitary equivalence class is not meager it will have to be comeager. So assume the unitary equivalence class of some T0 is comeager, towards a contradiction. Then for comeager many T ∈ Aut(X, µ), there is U ∈ U (L2 (X, µ)) with U UT U −1 = UT0 . By the Jankov-von Neumann Selection Theorem, we can find a Borel function f : Aut(X, µ) → U (L2 (X, µ)) so that, letting fT = f (T ), {(S, T ) : ∃n[kUSn (χA ) − χA k <
∀∗ T (fT UT fT−1 = UT0 ),
8
I. MEASURE PRESERVING AUTOMORPHISMS
where S ∀∗ T means “for comeager many T .” Fix a dense set {ξm } ⊆ L2 (X, µ). Then m {T : kfT (χA ) − ξm k < 1/20} = Aut(X, µ). So there is an open nonempty set W ⊆ Aut(X, µ) and some m so that, putting ξm = ξ, we have 1 ∗ ∀ T ∈ W kfT (χA ) − ξk < . 20 We can then find S, T ∈ W and n such that 1 1 kUSn (χA ) − χA k < , kUTn (χA ) − χA k > 20 3 −1 −1 and fS US fS = fT UT fT = UT0 , 1 1 , kfT (χA ) − ξk < . 20 20 Then if fS (χA ) = ξ1 , fT (χA ) = ξ2 , we have 1 1 kUTn0 (ξ1 ) − ξ1 k < , kUTn0 (ξ2 ) − ξ2 k > , 20 3 1 and kξ1 − ξk, kξ2 − ξk < 20 , so kfS (χA ) − ξk <
kUTn0 (ξ) − ξk ≤ kUTn0 (ξ) − UTn0 (ξ1 )k + kUTn0 (ξ1 ) − ξ1 k + kξ1 − ξk, which is smaller than kUTn0 (ξ) − ξk ≥ which is bigger than
3 20 , while also kUTn0 (ξ2 ) − ξ2 k
7 30 ,
− kUTn0 (ξ2 ) − UTn0 (ξ)k − kξ − ξ2 k,
a contradiction.
2
Remark. One can employ a similar strategy to give another proof of the well-known fact (see Nadkarni [Na], Ch. 8) that in the unitary group U (H) of a separable infinite dimensional Hilbert space H (with the weak, equivalently strong, topology) every conjugacy class is meager. Fix a unit vector v ∈ H. As in the proof of 2.5, it is enough to show that for any positive , δ < 1 the open set W = {(S, T ) ∈ U (H)2 : ∃n(kS n (v) − vk < and kT n (v) − vk > δ)} is dense in U (H)2 . Fixing a basis {bi }∞ i=1 for H, with v = b1 , we can view the unitary group U (m) S of Cm as the unitary group of the subspace hbi im i=1 and it is easy to see that m U (m) is dense in U (H). So it is enough to show that given any m and S0 , T0 ∈ U (m), there are S, T ∈ U (m) as close we want to S0 , T0 such that for some n, kS n − Ik < and kT n + Ik < 2 − δ, where I is the identity of U (m) and k·k refers here to the norm of operators on Cn . For this it is enough to show again that given any U ∈ U (m) and ρ > 0, there is N such that for n ≥ N , there are V, W ∈ U (m) with kU − V k, kU − W k < ρ and V n = I, W n = −I. Since any U0 ∈ U (m) is conjugate in U (m) to some U1 ∈ U (m) which is diagonal for an orthonormal basis of Cm , i.e., for such a basis {e1 , . . . em } we have U1 (ei ) = λi ei , where λi ∈ T, we can assume that U is already of that form. Given ρ > 0, there is σ > 0, such that if P ∈ U (m) satisfies P (ei ) = µi ei , where |µi − λi | < σ, 1 ≤ i ≤ m, then kU − P k < ρ. Now we
2. SOME BASIC FACTS ABOUT Aut(X, µ)
9
can find large enough N , such that if n ≥ N , there are nth roots of 1, say λ01 , . . . , λ0m , with |λ0i − λi | < σ, and nth roots of −1, say λ001 , . . . , λ00m , with |λ00i − λi | < σ. Then if V (ei ) = λ0i ei , W (ei ) = λ00i ei , V, W clearly work. Addendum. Rosendal [Ro] has recently found the following simple proof that all unitary equivalence classes in Aut(X, µ) are meager in the weak topology. For each infinite set I ⊆ N, let A(I) = {T ∈ Aut(X, µ) : ∃i ∈ I(T i = 1)}. By 2.2 and 2.3, A(I) is dense in (Aut(X, µ), w). Let V0 ⊇ V1 ⊇ . . . be a basis of open nbhds of 1 and consider the set B(I, k) = {T ∈ Aut(X, µ) : ∃i ∈ I(i > k & T i ∈ Vk }. It contains A(I \ {0, . . . , k}), so it is open dense. Thus \ C(I) = B(I, k) = {T : ∃{in } ⊆ I(T in → 1)} k
is comeager and clearly invariant under unitary equivalence. If a unitary equivalence class C is non-meager, it is contained in all C(I), I ⊆ N infinite. It follows that if T ∈ C, then T n → 1, so T = 1, a contradiction. Notice that this proof shows that in any Polish group G in which the sets A(I) = {g ∈ G : ∃i ∈ I(g i = 1)} are dense for all infinite I ⊆ N, every conjugacy class is meager. Another group that has this property is U (H) (see the preceding remark) and this gives a simple proof that conjugacy classes in U (H) are meager. Denote by ERG the set of ergodic T ∈ Aut(X, µ). Theorem 2.6 (Halmos [Ha]). ERG is dense Gδ in (Aut(X, µ), w). Proof. Take X = 2N . For T ∈ Aut(X, µ), define fT : X → X Z by fT (x)n = T −n (x). Then if S is the (left) shift action on X Z , S((xn )) = (xn−1 ), clearly fT (T n (x)) = S n (fT (x)). So if νT = (fT )∗ µ, νT is shift-invariant. We will verify that T 7→ νT is continuous from (Aut(X, µ), w) to the space of shift-invariant probability measures on X Z , which is a convex compact set in the usual weak∗ -topology of probability measures on the compact space X Z . Then T ∈ ERG ⇔ νT is an ergodic, shift-invariant measure. But the ergodic shift-invariant measures are exactly the extreme points of the convex, compact set of shift-invariant measures, so they form a Gδ set and thus ERG is Gδ . To check the continuity of T 7→ νT we needRto verify that for some uniformly dense D ⊆ C(X Z ), and any f ∈ D, T 7→ f dνT is continuous. By Stone-Weierstrass we can take D to be the complex algebra generated by the functions (xi )i∈Z 7→ χA (xn ), for n ∈ Z, A ⊆ X clopen. Thus it is enough
10
I. MEASURE PRESERVING AUTOMORPHISMS
to check that for each finite {n1 , . . . , nk } ⊆ Z and clopen sets {A1 , . . . , Ak } the function T 7→
Z Y k
χAi (xni )dνT ((xn ))
i=1 k Y
Z
χAi (T −ni (x))dµ(x)
=
i=1
=
Z Y k
χT ni (Ai ) (x)dµ(x)
i=1
Z =
χTk
i=1
= µ(
k \
T ni (Ai ) dµ
T ni (Ai ))
i=1
is continuous, which is clear.
2
Concerning stronger notions of ergodicity, the set of weak mixing transformations, WMIX, is also dense Gδ (Halmos [Ha]; see 12.1 for a more general statement) but the set of mild mixing transformations, MMIX, and the set of (strong) mixing transformations, MIX, are meager (Rokhlin for MIX; see Katok-Thouvenot [KTh], 5.48 and Nadkarni [Na], Chapter 8). (B) Recall that if T ∈ Aut(X, µ) and UT ∈ U (L2 (X, µ)) is the corresponding unitary operator, then ERG, MMIX, WMIX, MIX are characterized as Rfollows in terms of UT , where we put below L20 (X, µ) = {f ∈ L2 (X, µ) : f dµ = 0} = C⊥ (the orthogonal of the constant functions): T ∈ ERG ⇔ UT (f ) 6= f, ∀f ∈ L20 (X, µ) \ {0}, T ∈ WMIX ⇔ UT (f ) 6= λf, ∀λ ∈ T∀f ∈ L20 (X, µ) \ {0}, T ∈ MMIX ⇔ UTnk (f ) 6→ f, ∀nk → ∞, ∀f ∈ L20 (X, µ) \ {0}, T ∈ MIX ⇔ hU n (f ), f i → 0, ∀f ∈ L20 (X, µ). We have MIX ⊆ MMIX ⊆ WMIX ⊆ ERG and these are proper inclusions. (Note that if for any T ∈ Aut(X, µ) we denote by κ0T the Koopman representation of Z on L20 (X, µ) induced by UT , then T is in ERG, etc., iff κ0T is ergodic, etc., according to the definition in Appendix H.) Also denote by σT0 the maximal spectral type of UT |L20 (X, µ) (thus σT0 is uniquely defined up to measure equivalence); see Appendix F (where we take ∆ = Z, H = L20 (X, µ), π(n) = UTn |L20 (X, µ)). Note that σT0 can be chosen so that the map T 7→ σT0 is continuous from (Aut(X, µ), w) into P (T), the
2. SOME BASIC FACTS ABOUT Aut(X, µ)
11
space of probability measures on T with the weak∗ -topology. Then we have T ∈ ERG ⇔ σT0 ({1}) = 0, T ∈ WMIX ⇔ σT0 is non-atomic, T ∈ MMIX ⇔ σT0 ∈ D⊥ , T ∈ MIX ⇔ σT0 ∈ R, where R is the class of Rajchman measures in P (T), i.e., those measures µ ∈ P (T) such that µ ˆ(n) → 0 as |n| → ∞, D is the class of closed Dirichlet sets in T, i.e., the class of closed E ⊆ T such that there is nk → ∞ with z nk → 1 uniformly for z ∈ E, and finally D⊥ is the class of all µ ∈ P (T) that annihilate all Dirichlet sets. All the preceding characterizations in terms of UT , σT0 can be found, for example, in Queff´elec [Qu]. The reader should note that in Queffel´ec [Qu], III.21 one finds the characterization: T ∈ MMIX ⇔ σT0 ∈ D⊥ in the form T ∈ MMIX ⇔ σT0 ∈ LI , where LI is a certain class of measures defined there. A proof that LI = D⊥ can be found in Proposition 3, p. 212 and Proposition 9, p. 243 of Host-M´ela-Parreau [HMP]. For more about Dirichlet sets, see Lindahl–Poulsen [LP], Ch.1. For each σR ∈ P (T) which is symmetric (i.e., invariant under conjugation), σ ˆ (n) = z −n dσ(z) is a positive-definite real function and let Tσ be the Gaussian shift on RZ corresponding to σ ˆ (see Appendix C). It is known (see Cornfeld-Fomin-Sinai [CFS], 14.3, Theorem 1) that the maximal specP 1 ∗n ∗n = σ ∗· · ·∗σ is the n-fold σ , where σ tral type of UTσ |C⊥ is σ∞ = ∞ n=1 2n convolution of σ. We also have (see Cornfeld-Fomin-Sinai [CFS], 14.2) and Queff´elec [Qu], III. 21): σ is non-atomic ⇔ Tσ ∈ ERG ⇔ Tσ ∈ WMIX, σ ∈ D⊥ ⇔ Tσ ∈ MMIX, σ ∈ R ⇔ Tσ ∈ MIX. For the equivalence σ ∈ D⊥ ⇔ Tσ ∈ MMIX one needs to notice that σ ∈ D⊥ ⇒ σ∞ ∈ D⊥ , which follows from the fact that D⊥ is closed under convolution by any measure, since D is closed under translations (see Lindahl–Poulsen [LP], p.4). We already noted that ERG, WMIX are Gδ in (Aut(X, µ), w) and it is easy to see that MIX is Π03 . However we have the following result. Theorem 2.7 (Kaufman, Kechris). The set of mild mixing transformations in Aut(X, µ) is co-analytic but not Borel in the weak topology of Aut(X, µ). Proof. It is clear that MMIX is co-analytic. The map σ 7→ Tσ ∈ Aut(X, µ) is Borel and D⊥ restricted to symmetric measures is known to be non-Borel (Kechris-Lyons [KLy], Host-Louveau-Parreau). 2 Remark. Recall that a subset A of N \ {0} is an IP-set if there are p1 , p2 , · · · ∈ N \ {0} such that A = {pi1 + · · · + pik : i1 < · · · ik }. The filter
12
I. MEASURE PRESERVING AUTOMORPHISMS
IP∗ consists of all A ⊆ N \ {0} that intersect every IP-set (see Furstenberg [Fur]). It is shown in [Fur], 9.22, that T ∈ Aut(X, µ) is mild mixing iff for any two Borel sets A, B ⊆ X, µ(A ∩ T −n (B)) → µ(A)µ(B) in the filter IP*. It is clear that A, B can be restricted to any countable dense subset of the measure algebra. It follows that IP*, which is clearly a co-analytic set in the compact metric space of subsets of N \ {0}, is not Borel. In Kechris-Lyons [KLy] a canonical Π11 -rank on D⊥ is defined and used to prove, by a boundedness argument, the non-Borelness of D⊥ . It might be interesting to define a canonical Π11 -rank on MMIX, as this would probably give an interesting transfinite hierarchy of progressively “milder” mild mixing notions. One could even wonder whether mixing would occupy exactly the lowest level of this hierarchy. (C) We next show that Aut(X, µ) is contractible. Theorem 2.8 (Keane [Kea]). The group Aut(X, µ) is contractible in both the weak and the uniform topology. Proof. For each A ∈ MALGµ , T ∈ Aut(X, µ), define the induced transformation TA ∈ Aut(X, µ) as follows: Let B0 = X \ A, B1 = A ∩ T −1 (A), . . . , Bn = A ∩ T −1 (B0 ) ∩ · · · ∩ T −(n−1) (B0 ) ∩ T −n (A), n ≥ 2. The Poincar´e Recurrence Theorem guarantees that {Bn } is a partition of X (a.e.). Let TA (x) = T n (x), if x ∈ Bn . Then (T, A) 7→ TA is continuous from (Aut(X, µ), u) × MALGµ → (Aut(X, µ), u) and also from (Aut(X, µ), w) × MALGµ → (Aut(X, µ), w). Assume now X = [0, 1], µ = Lebesgue measure, and define ϕ : [0, 1] × Aut(X, µ) → Aut(X, µ) by ϕ(λ, T ) = T[λ,1] . Then this is a contraction to the identity of Aut(X, µ). 2 For the weak topology a complete topological characterization has been obtained. Theorem 2.9 (Nhu [Nh]). The space (Aut(X, µ), w) is homeomorphic to `2 . (D) Finally we mention a number of algebraic properties of Aut(X, µ). Theorem 2.10 (Fathi [Fa], Eigen [Ei2], Ryzhikov [Ry1]). The group Aut(X, µ) has the following properties: (i) Every element is a commutator and the product of 3 involutions. (ii) It is a simple group. (iii) Every automorphism is inner. Proofs of (ii), (iii) and a somewhat weaker version of (i) are contained in the arguments given in the course of the proof of Theorem 4.1 below Comments. For 2.5 see Nadkarni [Na], §8 and references contained therein. For 2.10, see Choksi-Prasad [CP] for some generalizations.
3. FULL GROUPS OF EQUIVALENCE RELATIONS
13
3. Full groups of equivalence relations (A) Let (X, µ) be a standard measure space and E a Borel equivalence relation on X. We say that E is a countable equivalence relation if every E-class is countable. Feldman-Moore [FM] have shown that E is countable iff there is a countable (discrete) group Γ acting in a Borel way on X such that if (γ, x) 7→ γ · x denotes the action and we let EΓX = {(x, y) : ∃γ ∈ Γ(γ · x = y)} be the induced equivalence relation, then E = EΓX . We say that E is a measure preserving equivalence relation if the action of Γ is measure preserving, i.e., x 7→ γ · x is measure preserving for each γ ∈ Γ. This condition is independent of Γ and the action that induces E and is equivalent to the condition that every Borel automorphism T of X for which T (x)Ex, ∀x, is measure preserving. For any Borel E we define its full group or inner automorphism group, [E], by [E] = {T ∈ Aut(X, µ) : T (x)Ex, µ-a.e.(x)}. It is easy to check that this is a closed subgroup of Aut(X, µ) in the uniform topology. Convention. Below we will assume that equivalence relations E are countable and measure preserving, unless otherwise explicitly stated. Recall that E is an ergodic equivalence relation if every E-invariant Borel set is null or co-null. We call E a finite equivalence relation or a periodic equivalence relation if (almost) all its classes are finite and an aperiodic equivalence relation if (almost) all its classes are infinite. Proposition 3.1. E is ergodic iff [E] is dense in (Aut(X, µ), w). Proof. Assume E is ergodic. Then if A, B are Borel sets of the same measure, there is T ∈ [E] with T (A) = B. Now let S ∈ Aut(X, µ). A nbhd basis for S is given by the sets of the form {T : ∀i ≤ n(dµ (T (Ai ), S(Ai )) < )}, where > 0, and A1 , . . . , An is a BorelSpartition of X. Find T1 , . . . , Tn ∈ [E] such that Ti (Ai ) = S(Ai ). Then T = ni=1 Ti |Ai is in [E] and clearly T (Ai ) = S(Ai ), i = 1, . . . , n. If E is not ergodic, there is A ∈ MALGµ with 0 < µ(A) < 1 which is [E]-invariant. If [E] was weakly dense in Aut(X, µ), A would be T -invariant for any T ∈ Aut(X, µ), which is absurd. 2 T.-J. Wei [We] has shown in fact that [E] is closed in w iff E is a finite equivalence relation. Otherwise, it is Π03 -complete in w and its closure in the weak topology is equal to [F ], where F is the (not necessarily countable) equivalence relation corresponding to the ergodic decomposition of E. Also clearly [E] is a meager subset of (Aut(X, µ), w), since it is a Borel non-open subgroup of (Aut(X, µ), w). Note now the following basic fact about the uniform topology of [E].
14
I. MEASURE PRESERVING AUTOMORPHISMS
Proposition 3.2. In the uniform topology, [E] is separable, thus ([E], u) is Polish. Proof. Since E is a countable equivalence relation, there is a countable group Γ ≤ Aut(X, µ) such that E = EΓX = the equivalence relation induced ∞ by Γ. Thus T ∈ [E] iff there is a Borel partition {An }∞ n=1 of X and {γn }n=1 ⊆ Γ such that T |An = γn |An . Fix also a countable Boolean algebra A of Borel sets closed under the Γ-action and dense in the measure algebra. Given C1 , . . . , CN ∈ A and γ1 , . . . , γN ∈ Γ such that C1 , . . . , CN are pairwise 1 (C1 ), . . . , γN (CN ) are pairwise disjoint, let C0 = S disjoint and γS X \ i≤N Ci , D0 = X \ i≤N γi (Ci ), so that µ(C0 ) = D0 , and pick TC0 ,D0 ∈ S Aut(X, µ) with T (C0 ) = D0 . Put finally SC,~ ~ γ = TC0 ,D0 |C0 ∪ i≤N γi |Ci . Clearly there are only countably many SC,~ . We will show that for any ~ γ ~ ~γ such that δu (T, S ~ ) < . It follows then T ∈ [E] and > 0 there are C, C,~γ
that [E] is separable (being contained in the uniform closure of the SC,~ ~ γ ). ∞ ∞ So fix T ∈ [E], > 0, a Borel partition {An }n=1 of X and {γn }n=1 ⊆ Γ such that T |An = γn |A Pn , ∀n ≥ 1. For each δ1 , δ2 > 0, choose N = N (δ1 ) large enough so that i>N µ(Ai ) < δ1 , and then B1 , . . . , BN ∈ A with µ(Ai ∆Bi ) < δ2 . Let for 1 ≤ i ≤ N , [ [ Ci = {x ∈ Bi : x 6∈ Bj & γi (x) 6∈ γj (Cj )}, j
j
so that {Ci }i≤N ⊆ A are pairwise disjoint and so are {γi (Ci )}i≤N ⊆ A. Then if δ1 , δ2 are chosen small enough, δu (T, SC,~ 2 ~ γ ) < . Since the identity is a continuous map from the Polish group ([E], u) onto the Borel subgroup [E] of (Aut(X, µ), w), it follows that [E] is a Polishable subgroup of (Aut(X, µ), w) with the unique corresponding Polish topology being the uniform topology. (B) Some of the basic facts we mentioned earlier for Aut(X, µ) “localize” to [E]. First the proof of Rokhlin’s Lemma actually shows that the Uniform Approximation Theorem is true in each [E]. Theorem 3.3 (Uniform Approximation Theorem for [E]). If T ∈ [E] is aperiodic, then for each N ≥ 1, > 0 there is a periodic S ∈ [E] of period N such that δu (S, T ) ≤ N1 + . From this we also have the following, in case E is also ergodic. Theorem 3.4 (Conjugacy Lemma for [E]). If E is ergodic, and T ∈ [E] is aperiodic, {ST S −1 : S ∈ [E]} is uniformly dense in APER ∩ [E] (which is a closed subset of ([E], u)). Proof. In the proof of 2.4 note that, since E is ergodic, there is S0 ∈ [E] with S0 (A) = B. 2 We also note another well-known basic fact.
3. FULL GROUPS OF EQUIVALENCE RELATIONS
15
Theorem 3.5. If E is ergodic (resp., aperiodic), then there is T ∈ [E] which is ergodic (resp., aperiodic). Proof. (i) We first give the proof for the aperiodic case, which works in the pure Borel context (with no measures present). The proof below is due to Dougherty. Let E be a countable aperiodic Borel equivalence relation on a standard Borel space X. Let < be a Borel linear order on X and let a countable group Γ act in a Borel way on X so that E = EΓX . Put Γ = {γn }. Let A = {x : [x]E has a largest element in <}. Then E|A admits a Borel transversal, so clearly there is an aperiodic Borel automorphism of A, say TA , with TA (x)Ex, ∀x ∈ A. Let now B = X \ A. Define U : B → B by U (x) = γn (x), where n is least with γn (x) > x. Consider then the equivalence relation on B: xEU y ⇔ ∃m∃n(U n (x) = U m (y)). There is a Borel automorphism of B, say TB , that induces EU (see DoughertyJackson-Kechris [DJK], 8.2). Clearly TB (x)Ex, ∀x ∈ B, and TB is aperiodic. So T = TA ∪ TB works. (ii) For the ergodic case, see Zimmer [Zi1], 9.3.2 or Kechris [Kec1], 5.66. Alternatively, one can go over the proof of Dye’s Theorem, 3.13 below, given in Kechris-Miller [KM], 7.13, and notice that if condition (4) is omitted in that argument, one obtains a proof that every ergodic E contains an ergodic subequivalence relation isomorphic to E0 , defined before 3.8 below, which is thus generated by an ergodic automorphism. 2 As an immediate consequence, we obtain the next result. Theorem 3.6. If E is ergodic, then ERG∩[E] is dense Gδ in (APER ∩ [E], u). Proof. Since ERG is Gδ in w ⊆ u, it is clearly Gδ in u. Density follows from the two preceding results. 2 Similarly (using also 3.14 below) one can show the same fact for the set of weak mixing T ∈ [E]. It will be useful, given a countable group Γ ≤ Aut(X, µ) and letting E = EΓX = {(x, y) : ∃γ ∈ Γ(γ(x) = y)}, to have a criterion for a countable subgroup ∆ ≤ [E] to be dense in ([E], u). Note that if ∆ ≤ [E] is dense X . (However, if E = E X , Γ may not be dense in in ([E], u), then E = E∆ Γ ([E], u).) Proposition 3.7. Suppose Γ ≤ Aut(X, µ) is a countable group and E = EΓX . Let also A ⊆ MALGµ be a countable Boolean algebra dense in MALGµ and closed under Γ. Let ∆ ≤ [E] be a countable subgroup of [E]. Then ∆ is dense in ([E], u) provided that for every > 0 and every A0 , . . . , An−1 ∈ n−1 A, γ0 , . . . , γn−1 ∈ Γ with {Ai }i=0 , {γi (An )}n−1 i=0 each pairwise disjoint, there is δ ∈ ∆ such that µ({x ∈ Ai : γi (x) 6= δ(x)}) < , i = 0, . . . , n − 1.
16
I. MEASURE PRESERVING AUTOMORPHISMS
The proof is similar to that of 3.2, so we omit it. Notice that this condition is also necessary if E is ergodic. This has the following consequence; below E0 denotes the following equivalence relation on 2N : xE0 y ⇔ ∃m∀n ≥ m(xn = yn ). Proposition 3.8. The group of dyadic permutations of 2N is uniformly dense in [E0 ]. Proof. We use 3.7. Take Γ = {dyadic permutations}, A = the algebra of clopen sets in 2N . Thus A consists of all sets of the form [ Ns , I ⊆ 2n for some n. s∈I N
Then EΓ2 = E0 and A is closed under Γ. To verify 3.7 for ∆ = Γ take {Ai }i
γi (Ai ) =
[
Ns , Ji ⊆ 2N ,
s∈Ji
and γi is a dyadic permutation of rank N , i = 0, . . . , n − 1. Thus each of {Ii }i
3. FULL GROUPS OF EQUIVALENCE RELATIONS
17
Otherwise, find (S0 , T0 ) ∈ (APER ∩ [E])2 , > 0 such that for every (S, T ) ∈ (APER ∩ [E])2 with δu (S, S0 ) < , δu (T, T0 ) < , we have w(S, T ) = 1, i.e., w(S, T )(x) = x, µ−a.e. (x). Let w1 , w2 , . . . , wn = w be the co-initial (non-∅) subwords of w (e.g., if w = xy 2 x−1 , w1 = x−1 , w2 = yx−1 , w3 = y 2 x−1 , w4 = xy 2 u−1 = w). Clearly n ≥ 2. Lemma 3.10. If P1 , . . . , Pk ∈ Aut(X, µ), and on a Borel set of positive measure A and every x ∈ A we have that P1 (x), P2 P1 (x), . . . , Pk · · · P1 (x) are distinct, then there is B ⊆ A of positive measure such that the sets P1 (B), P2 P1 (B), . . . , Pk · · · P1 (B) are pairwise disjoint. Proof. We can assume that X is a Polish space, A ⊆ X is clopen and P1 , . . . , Pk are homeomorphisms. Fix a countable basis {Un } for X. Then every point x ∈ A is contained in a basic open set Un(x) S such that P1 (Un(x) ), . . . , Pk · · · P1 (Un(x) ) are pairwise disjoint. As A ⊆ {Un(x) : x ∈ A}, for some x ∈ A, B = Un(x) ∩ A has positive measure. 2 For each T ∈ Aut(X, µ), let ET be the equivalence relation induced by T, xET y ⇔ ∃n ∈ Z(T n (x) = y) and let [T ] = [ET ]. Lemma 3.11. Given aperiodic T ∈ Aut(X, µ) and a Borel set with µ(A) > 0, there is a Borel set A0 ⊆ A with µ(A0 ) > 0 and aperiodic T 0 ∈ [T ] such that T (x) = T 0 (x), µ-a.e. x 6∈ A0 , T (x) 6= T 0 (x), µ-a.e. x ∈ A0 . Proof. We can assume that A intersects every orbit of T and, by Poincar´e recurrence, that for (almost) every x ∈ A there are k, ` > 0 with T k (x), T −` (x) ∈ A. We can also assume that if B = X \ A, then µ(B) > 0 and ∀x ∈ B∃k, ` > 0(T k (x) ∈ B, T −` (x) ∈ B) (since if µ(B) = 0 we can take A0 = A = X, T 0 (x) = T 2 (x)). Then we can find A0 ⊆ A ∩ [B]T , where [B]T denotes the T -saturation of B, such that [A0 ]T = [B]T , so that µ(A0 ) > 0, ∀x ∈ A0 ∃k > 0∃` > 0(T k (x), T −` (x) ∈ A0 ) and finally x ∈ A0 ⇒ T (x) 6∈ A0 . Then we put T 0 (x) = T (x) if x 6∈ A0 , and T 0 (x) = T (T k (x)), where k > 0 is least with T k (x) ∈ A0 , if x ∈ A0 . 2 For convenience, in the rest of the argument, we will write, for any reduced word v(x, y), v S,T instead of v(S, T ). Continuing the proof, let 0 < k0 ≤ n be least such that on a set of positive measure A0 we have ,T0 x ∈ A0 ⇒ x, w1S0 ,T0 (x), . . . , wkS00−1 (x0 ) are distinct and ,T0 wkS00 ,T0 (x) ∈ {x, . . . , wkS00−1 (x)}.
Clearly k0 ≥ 2 and actually ,T0 wkS00 ,T0 (x) ∈ {x, . . . , wkS00−2 (x)}
18
I. MEASURE PRESERVING AUTOMORPHISMS
(since S0 , T0 are aperiodic). By 3.10, we can also assume that ,T0 (A0 ) A0 , w1S0 ,T0 (A0 ), . . . , wkS00−1
are pairwise disjoint and if the first symbol of wk0 is say S0±1 (the other case ,T0 (A0 )) = wiS0 ,T0 (A0 ), for some 0 ≤ i ≤ k0 − 2. being similar), then S0±1 (wkS00−1 Finally, we can assume that µ(A0 ) < /2. If the first symbol of wk0 is S0 (the other case being similar), then by 3.11, applied to the set A = wiS0 ,T0 (A0 ) and T = S0−1 , we can find the appropriate T 0 , A0 ⊆ A and then, letting S00 = (T 0 )−1 , A00 = (wiS0 ,T0 )−1 (A0 ) we have an aperiodic S00 with δu (S0 , S00 ) < /2 and A00 ⊆ A0 , µ(A00 ) > 0 such that S 0 ,T S 0 ,T S 0 ,T x, w1 0 0 (x), . . . , wk00−10 (x), wk00 0 (x) S 0 ,T0
are distinct for x ∈ A00 . (Note here that ∀j < k0 , wj 0 S 0 ,T
(x) = wjS0 ,T0 (x), ∀x ∈
A00 .) Since δu (S0 , S00 ) < /2, we still have wn0 0 (x) = x for almost all x ∈ A00 . Then, if k0 < n, let 0 < k0 < k1 ≤ n be least such that on a set A1 ⊆ A00 of measure < /4 S 0 ,T0
x ∈ A1 ⇒ x, w1 0
S 0 ,T0
wk10
S 0 ,T
(x), . . . , wk10−10 (x) are distinct and S 0 ,T0
(x) ∈ {x, w1 0
S 0 ,T
(x), . . . , wk10−10 (x)}
¯ T¯ with and repeat this process finitely many times to eventually find S, ¯ S0 ) < , δu (T¯, T0 ) < and a set A¯ of positive measure such that δu (S, ¯ ¯ ¯ ¯ x ∈ A¯ ⇒ x, w1S,T (x), . . . , wnS,T (x) are distinct, ¯ ¯
¯ wnS,T (x) = x. which is a contradiction, as for almost all x ∈ A,
2
A similar argument shows that the set of (S1 , . . . , Sn ) ∈ (APER ∩ [E])n that generate a free group is also dense Gδ . So using also the Mycielski, Kuratowski Theorem (see Kechris [Kec2], 19.1) this shows that there is a Cantor set C ⊆ APER ∩ [E] generating a free group, so [E] contains a free subgroup with continuum many generators. (D) Note that the proof of 2.8 also shows the following. Theorem 3.12 (Keane). The full group [E] is contractible for both the weak and the uniform topologies. Recently Kittrell and Tsankov [KiT] have shown that in fact ([E], u) is homeomorphic to `2 . (E) Recall that E is a hyperfinite equivalence relation if it is induced by an element of Aut(X, µ). In particular E0 is hyperfinite. The fundamental result about hyperfinite equivalence relations is the following classical theorem of Dye. For a proof see, e.g., Kechris-Miller [KM].
3. FULL GROUPS OF EQUIVALENCE RELATIONS
19
Theorem 3.13 (Dye). Let E, F be ergodic hyperfinite Borel equivalence relations. Then there is T ∈ Aut(X, µ) such that for all x, y in a set of measure 1, xEy ⇔ T (x)F T (y). Since the hyperfinite E which preserve µ and are ergodic are exactly the ones induced by ergodic T ∈ Aut(X, µ), we have for any two ergodic T1 , T2 ∈ Aut(X, µ), that [T1 ], [T2 ] are conjugate in Aut(X, µ). Also from 3.5 we have the next result. Corollary 3.14. If E is ergodic, then for any ergodic T ∈ Aut(X, µ), [E] contains a conjugate of T , i.e., [E] meets every conjugacy class in ERG. Thus, up to conjugacy, there is only one full group of a hyperfinite ergodic equivalence relation and it is the smallest (in terms of inclusion) among all full groups of ergodic countable equivalence relations. (F) We conclude with an application of 3.8 due to Giordano-Pestov. Recall that a topological group G is extremely amenable if every continuous action of G on a (Hausdorff) compact space has a fixed point. Theorem 3.15 (Giordano-Pestov [GP]). Let E be an ergodic, hyperfinite equivalence relation. Then ([E], u) is extremely amenable. Corollary 3.16 (Giordano-Pestov [GP]). The group (Aut(X, µ), w) is extremely amenable. The corollary follows from 3.15, since, by 3.1, the identity map is a continuous embedding of ([E], u) into (Aut(X, µ), w) with dense range. Proof of 3.15. We take E = E0 . Denote by Gn = S2n the finite group of permutations of 2n , which we identify with a subgroup of [E], identifying π ∈ S2n with S the element of [E] given by sˆx 7→ π(s)ˆx. Clearly G1 ⊆ G2 ⊆ . . . and n Gn is dense in ([E0 ], u) by 3.8. Now the metric δu restricted to Gn is clearly the Hamming metric on Gn : δu (π, ρ) = 21n card{s ∈ 2n : π(s) 6= ρ(s)}, and therefore, by a result of Maurey [Ma], the family (Gn , δu |Gn , µn ), where µn = counting measure on Gn , is a L´evy family and so, by Gromov-Milman [GM], ([E0 ], u) is a L´evy group and thus extremely amenable (see, e.g., Pestov [Pe2]). 2 Giordano-Pestov [GP] also prove that, conversely, if for an ergodic E the group ([E], u) is extremely amenable, then E is hyperfinite, so this gives a nice characterization of the hyperfiniteness of E in terms of properties of [E]: Theorem 3.17 (Giordano-Pestov [GP]). If E is an ergodic equivalence relation, then E is hyperfinite iff ([E], u) is extremely amenable.
20
I. MEASURE PRESERVING AUTOMORPHISMS
4. The Reconstruction Theorem (A) We will prove here another beautiful result of Dye that asserts that any ergodic E is completely determined by [E] as an abstract group. Given two countable Borel measure preserving equivalence relations E, F we say that E, F are isomorphic, in symbols E∼ = F, if there is T ∈ Aut(X, µ) such that except on a null set xEy ⇔ T (x)F T (y), or, equivalently, T [E]T −1 = [F ], i.e., [E], [F ] are conjugate in Aut(X, µ). Theorem 4.1 (Dye). Let E, F be ergodic equivalence relations. Then the following are equivalent: (i) E ∼ = F. (ii) [E], [F ] are conjugate in Aut(X, µ). (iii) ([E], u), ([F ], u) are isomorphic as topological groups. (iv) [E], [F ] are isomorphic as abstract groups. Moreover, for any algebraic isomorphism f : [E] → [F ], there is unique ϕ ∈ Aut(X, µ) with f (T ) = ϕT ϕ−1 , ∀T ∈ [E]. (B) In preparation for the proof, we prove some results that are also useful for other purposes. Lemma 4.2. Let E be ergodic. Then every element of [E] is a product of 5 commutators in [E]. Proof. We will need two sublemmas. Sublemma 4.3. Let E be ergodic and let T ∈ [E] be periodic. Then T is a commutator in [E] and the product of two involutions in [E]. Proof. We can assume that for some n ≥ 2 all T -orbits have exactly n elements. Let A be a Borel selector for the T -orbits. Split A = A1 ∪ A2 , where µ(A1 ) = µ(A2 ) = 21 µ(A). Let Q ∈ [E] be an involution that sends S Sn−1 k k −1 |B . B1 = n−1 2 k=0 T (A1 ) to B2 = k=0 T (A2 ) and conjugates T |B1 to T −1 Let T1 = T |B1 ∪ id|B2 , T2 = id|B1 ∪ T |B2 . Then QT2 Q = T1 and T = T1 T2 = QT2−1 QT2 = [Q, T2−1 ] = Q(T2−1 QT2 ). 2 Sublemma 4.4. Let E be ergodic. If T ∈ [E], then we can write T = ST 0 , where S ∈ [E], S = [U, V ] = U1 V1 with U, V, U1 , V1 ∈ [E], δu (T 0 , 1) ≤ 1 2 δu (T, 1), U1 , V1 involutions and supp(U ) ∪ supp(V ) ∪ supp(U1 ) ∪ supp(V1 ) ∪ supp(T 0 ) ⊆ supp(T ). Proof. We can assume that T is aperiodic (since on the periodic part of T we can apply 4.3 and take T 0 = 1). We can then find a Borel complete section A for the T -orbits such that in the usual Z-order of each T -orbit the distance between successive elements of A is at least 2, and A is unbounded in each direction in every orbit (see, e.g., [KM], 6.7). Let T 0 = TA be the
4. THE RECONSTRUCTION THEOREM
21
induced transformation. So T 0 ∈ [E] and T (T 0 )−1 is periodic, thus, by 4.3, S = T (T 0 )−1 = [U, V ] = U1 V1 , for some U, V, U1 , V1 ∈ [E], U1 , V1 involutions. Also supp(T 0 ) = A and µ(A) ≤ 21 µ(supp(T )) = 12 , as A ∩ T (A) = ∅, so δu (T 0 , 1) ≤ 12 . 2 Fix now T ∈ [E]. All transformations below are in [E]. Then, by 4.4, T = S0 T1 , where S0 is a commutator and δu (T1 , 1) = µ(supp(T1 )) ≤ 12 . We concentrate on T1 . Split X = X1 ∪ X2 ∪ . . . , with µ(Xi ) = 2−i , supp(T1 ) ⊆ X1 . Then, by 4.4 again, T1 = S1 T20 , where S1 is a commutator of elements with support contained in X1 , T20 has support also contained in X1 , and δu (T20 , 1) ≤ 14 . Then find an involution U1,2 supported by X1 ∪ X2 , with U1,2 (supp(T20 )) ⊆ X2 , thus, as in general W (supp(V )) = supp(W V W −1 ), we have supp(U1,2 T20 U1,2 ) ⊆ X2 . Let T2 = U1,2 T20 U1,2 , so that T1 = S1 U1,2 T2 U1,2 = S1 (U1,2 T2 U1,2 T2−1 )T2 = S1 S1,2 T2 , where S1,2 is a commutator of elements supported by X1 ∪ X2 . Continuing this way, we write T2 = S2 S2,3 T3 , where S2 is a commutator of elements supported by X2 and S2,3 a commutator of elements supported in X2 ∪X3 and T3 is supported by X3 , etc. Then if S¯n = Sn Sn,n+1 , S¯n is the product of two commutators supported by Xn ∪ Xn+1 , T1 = S¯1 T2 , T2 = S¯2 T3 , . . . , with Tn supported by Xn . Let P1 be equal to S¯2n+1 , n ≥ 0, on X2n+1 ∪ X2n+2 and P2 be equal to S¯2n+2 , n ≥ 0, on X2n+2 ∪ X2n+3 and the identity otherwise. Then P1 = S¯1 S¯3 S¯5 . . . = (T1 T −1 )(T3 T −1 )(T5 T −1 ) . . . 6 −1 −1 −1 (T1 T3 T5 . . . )(T2 T4 T6 . . . ) 2
=
4
and P2 = S¯2 S¯4 S¯6 . . . = (T2 T3−1 )(T4 T5−1 )(T6 T7−1 ) . . . = (T2 T4 T6 . . . )(T3−1 T5−1 T7−1 . . . ). Thus T1 = P1 P2 and P1 , P2 are each a product of 2 commutators.
2
Lemma 4.5. Let E be ergodic. Then every element of [E] is a product of 10 involutions in [E]. Proof. Let T ∈ [E]. All transformations below are in [E]. As in the preceding proof, we write T = S0 T1 , where δu (T1 , 1) ≤ 21 and, by 4.4, S0 a product of 2 involutions. Then, proceeding as before, we have T1 = S1 T20 = S1 (T20 U1,2 (T20 )−1 )U1,2 (U1,2 T20 U1,2 ), with S1 a product of 2 involutions, and 0 T , where S 0 = note that U1,2 , T20 U1,2 (T20 )−1 are involutions. So T1 = S1 S1,2 2 1,2 (T20 U1,2 (T20 )−1 )U1,2 is the product of two involutions. Thus, continuing as in the rest of the proof, T1 is the product of two transformations each of which is the product of 4 involutions and we are done. 2 We now have as a consequence of 4.2 the following result. Theorem 4.6. If E is ergodic, then [E] is a simple group. Proof. Let N [E], N 6= {1}. We will show that N = [E]. Notice that for any group G and N G, if H generates G and [H, H] ⊆ N , then also
22
I. MEASURE PRESERVING AUTOMORPHISMS
[G, G] ⊆ N . Indeed if π : G → G/N is the canonical projection, every two elements of π(H) commute, so since π(H) generates π(G), π(G) is abelian, so [G, G] ⊆ N . Also notice that for each > 0, the elements T ∈ [E] with δu (T, 1) < generate [E]. This follows immediately from 4.5 and the fact that any involution in [E] is a product of involutions in [E] with arbitrarily small support (or alternatively one can use the path connectedness of [E], see 3.12). Finally fix T0 6= 1 in N and then a set A of positive measure with T0 (A) ∩ A = ∅. Let = µ(A)/2. We will show that the commutator of any two elements of [E] with δu (T, 1) < is in N . Fix S, T ∈ [E] with δu (T, 1), δu (S, 1) < . Conjugating, if necessary, we can assume that their supports are contained in A. Thus S −1 , T0 T −1 T0−1 have disjoint supports, so commute. Consider Tˆ = (T T0 T −1 )T0−1 ∈ N . Then [S, T ] = ST S −1 T −1 = ST (T0 T −1 T0−1 )S −1 (T0 T T0−1 T −1 ) = S(T T0 T −1 T0−1 )S −1 (T0 T T0−1 T −1 ) = S TˆS −1 (Tˆ)−1 ∈ N. 2 Note that conversely simplicity of [E] implies that E is ergodic. Indeed, if A is an E-invariant Borel set which is neither null or conull, then the elements of [E] supported by A form a non-trivial closed normal subgroup of [E]. In fact, Bezuglyi-Golodets [BG1] have shown that in case E is aperiodic, non-ergodic, all the nontrivial closed normal subgroups of [E] arise this way by letting A vary over the invariant Borel sets. We need one more lemma before embarking in the proof of 4.1. Lemma 4.7. Let E be ergodic. Let 1 6= T ∈ [E] be an involution, CT its centralizer in [E]. Then [T ] is the largest abelian normal subgroup of CT . Proof. Note that for S, T 0 ∈ Aut(X, µ), S[T 0 ]S −1 = [ST 0 S −1 ]. So if S ∈ CT , S[T ]S −1 = [T ], i.e., [T ] CT . Clearly, as T is an involution, [T ] is abelian. Let now N CT , N abelian. Consider the standard Borel space X/T = X/ET , the projection π : X → X/T and the measure ν = π∗ µ. Then E/T is a countable, measure preserving, ergodic equivalence relation on (X/T, ν). Clearly every S ∈ CT preserves the support of T and induces a map S ∗ in the full group of E/T restricted to supp(T )/T ; ρ(S) = S ∗ is clearly a surjective homomorphism. So ρ(N ) is a normal subgroup of [(E/T )|(supp(T )/T )], which is simple, by 4.6. Since ρ(N ) is abelian, ρ(N ) 6= [(E/T )|(supp(T )/T )], thus ρ(N ) = {1}, i.e., N ⊆ ker(ρ) = CT ∩ [T ] ⊆ [T ], which completes the proof. 2 (C) We are now ready to give the proof of 4.1. It is clearly enough to show (iv) ⇒ (i) and the “moreover” assertion.
4. THE RECONSTRUCTION THEOREM
23
So fix an algebraic isomorphism f : [E] → [F ]. It follows from 4.7 that if T ∈ [E], 1 6= T, T 2 = 1, then f ([T ]) = [f (T )]. Let for T ∈ [E] (and similarly for [F ]), ⊥(T ) = {S ∈ [E] : supp(S) ∩ supp(T ) = ∅}. Lemma 4.8. If T 6= 1 is an involution, then f (⊥(T )) = ⊥(f (T )). Proof. Note that ⊥(T ) is really the same as [E|(X \ supp(T ))], so every element of ⊥(T ) is the product of commutators in ⊥(T ), by 4.2. Thus the same holds for f (⊥(T )). We will show that f (⊥(T )) ⊆ ⊥(f (T )). Applying this to f −1 , f (T ) we get f −1 (⊥(f (T ))) ⊆ ⊥(T ), so we have equality. Claim. If S ∈ C[T ] , where C[T ] is the centralizer of [T ] in [E], then S = S1 S2 , where S1 ∈ [T ], S2 ∈ ⊥(T ). Granting this claim, we complete the proof of 4.8. Take any commutator in f (⊥(T )), say [U1 , U2 ], with U1 , U2 ∈ f (⊥(T )). Then, as ⊥(T ) ⊆ C[T ] , Ui ∈ f (C[T ] ) = C[f (T )] (this is because [T ] and thus C[T ] is preserved under f , by 4.7). So, by the claim, applied to f (T ), Ui = Ti Vi with Ti ∈ [f (T )], Vi ∈ ⊥(f (T )). Thus U1 U2 U1−1 U2−1 = T1 V1 T2 V2 V1−1 T1−1 V2−1 T2−1 and, since Ti , Vj commute, this is equal to V1 V2 V1−1 V2−1 ∈ ⊥(f (T )). Proof of the claim. If S ∈ C[T ] , then S ∈ CT , so S(supp(T )) = supp(T ), thus S = S1 S2 , where S1 = S|supp(T ) ∪ id|(X \ supp(T )), S2 = S|(X \ supp(T )) ∪ id|supp(T ). Clearly S2 ∈ ⊥(T ), so it is enough to show that S1 ∈ [T ]. By restricting everything to the support of T , if necessary, it is enough to show that if supp(T ) = X and S ∈ C[T ] , then S ∈ [T ]. Now S descends to S ∗ on X/T , so if S 6∈ [T ], S ∗ 6= 1 on X/T , thus there is a T -invariant set A ⊆ X of positive measure with S(A) ∩ A = ∅. Then S(A) is also T -invariant. Define now U ∈ [T ] to be the identity on X \ S(A) and switch any two elements of S(A) in the same T -orbit. Then U S 6= SU , a contradiction. 2 Now take any A ∈ MALGµ . Then there is an involution T in [E] with supp(T ) = A (split A into two sets of equal measure and let T map one onto the other and be the identity off A). Put Φ(A) = supp(f (T )) ∈ MALGµ . Claim. This is well-defined. Proof of the claim. Indeed, if T 0 is another involution with supp(T 0 ) = A, we have ⊥(T ) = ⊥(T 0 ), so ⊥(f (T )) = ⊥(f (T 0 )), by 4.8. Note now that for any U, V ∈ [F ], supp(U ) ⊆ supp(V ) ⇔ ⊥(V ) ⊆ ⊥(U ), so supp(U ) = supp(V ) ⇔ ⊥(V ) = ⊥(U ), thus supp(f (T )) = supp(f (T 0 )). 2
24
I. MEASURE PRESERVING AUTOMORPHISMS
Clearly Φ is a bijection of MALGµ with itself with inverse Φ−1 (B) = supp(f −1 (T 0 )), where T 0 is an involution with support B. Also if supp(T ) = A, supp(T 0 ) = B, A ⊆ B ⇔ supp(T ) ⊆ supp(T 0 ) ⇔ ⊥(T 0 ) ⊆ ⊥(T ) ⇔ ⊥(f (T 0 )) ⊆ ⊥(f (T )) ⇔ supp(f (T )) ⊆ supp(f (T 0 )) ⇔ Φ(A) ⊆ Φ(B). Thus Φ is an automorphism of the Boolean algebra MALGµ (not necessarily the measure algebra MALGµ ), so it is given by a (unique modulo null sets) Borel automorphism ϕ : X → X which is non-singular, i.e., preserves null sets, via Φ(A) = ϕ(A). We now want to verify that actually ϕ is measure preserving, i.e., ϕ ∈ Aut(X, µ), and moreover that f (S) = ϕSϕ−1 , ∀S ∈ [E]. Lemma 4.9. For each S ∈ [E], f (S) = ϕSϕ−1 . Granting this we verify that ϕ ∈ Aut(X, µ).R Since ϕ is non-singular, there is measurable r : X → R+ with µ(ϕ(A)) = A rdµ, ∀A ∈ MALGµ . We will show that r is actually constant, r = k. Then µ(ϕ(A)) = kµ(A), so, putting A = X, k = 1, thus ϕ ∈ Aut(X, µ). If r is not constant, there are two disjoint sets of positive measure A, B with 0 ≤ r(x) < λ < r(y), for x ∈ A, y ∈ B. Let C ⊆ A, D ⊆ B be such that µ(C) = µ(D) > 0 and let R K = ϕ(C). Let also U ∈ [E] be such that U (C) = D. Then µ(ϕ(D)) = D rdµ > λµ(D) = λµ(C), µ(K) = µ(ϕ(C)) = R −1 C rdµ < λµ(C) < µ(ϕ(D)). So µ(K) < µ(ϕ(D)) and ϕ(D) = ϕU ϕ (K), −1 −1 so µ(K) < µ(ϕU ϕ (K)), thus ϕU ϕ 6∈ Aut(X, µ), contradicting that ϕU ϕ−1 = f (U ) ∈ [F ] ⊆ Aut(X, µ). Proof of Lemma 4.9. As [E] is generated by involutions, it is enough to prove this for S = T , an involution. Notice that if T ∈ [E] is an involution, then supp(ϕT ϕ−1 ) = ϕ(supp(T )) = Φ(supp(T )) = supp(f (T )), i.e., ϕT ϕ−1 and f (T ) have the same support A. Assume, towards a contradiction, that B = {x ∈ A : ϕT ϕ−1 (x) 6= f (T )(x)}
4. THE RECONSTRUCTION THEOREM
25
has positive measure. Then we can find C ⊆ B of positive measure such that ϕT ϕ−1 (C) ∩ C = ∅, ϕT ϕ−1 (C) ∩ f (T )(C) = ∅, f (T )(C) ∩ C = ∅. Let D = C ∪ f (T )(C). Then D is f (T )-invariant but not ϕT ϕ−1 invariant. Write f (T ) = U1 U2 , where U1 = f (T ) on D, = id on X \ D and U2 = f (T ) on X \ D, id on D, so that U1 , U2 ∈ [F ] are involutions. Then T = f −1 (U1 )f −1 (U2 ), so ϕT ϕ−1 = (ϕf −1 (U1 )ϕ−1 )(ϕf −1 (U2 )ϕ−1 ). Now ϕf −1 (U1 )ϕ−1 has support ϕ(supp(f −1 (U1 ))) = supp(f f −1 (U1 )) = supp(U1 ) and similarly the automorphism ϕf −1 (U2 )ϕ−1 has support equal to supp(U2 ), thus ϕT ϕ−1 leaves D invariant, a contradiction. 2 Finally, we verify that there is unique ϕ such that f (S) = ϕSϕ−1 . For that we use the following lemmas which, for further reference, we state in more generality than we need here. Lemma 4.10. For any aperiodic E (not necessarily ergodic) and Borel set A, there is T ∈ [E] with supp(T ) = A. Proof. Since E is aperiodic, measure preserving, we can assume that E|A is aperiodic. Thus, by 3.5, there is aperiodic T0 : A → A with T0 (x)Ex. Let T = T0 ∪ id|(X \ A). 2 Corollary 4.11. For any aperiodic E (not necessarily ergodic) the centralizer C[E] of [E] in Aut(X, µ) is trivial. Proof. If S ∈ C[E] and A ∈ MALGµ , then by 4.10 there is T ∈ [E] with supp(T ) = A. Then A = supp(T ) = supp(ST S −1 ) = S(A), so S = 1. 2 The uniqueness now follows immediately from 4.11 and the proof of 4.1 is complete. Remark. Note that the argument following the statement of 4.9 shows that if ϕ is a non-singular Borel automorphism of X, E is ergodic and ϕ[E]ϕ−1 ⊆ Aut(X, µ), then ϕ ∈ Aut(X, µ). (D) Dye’s Reconstruction Theorem suggests the problem of distinguishing, up to isomorphism, ergodic equivalence relations by finding algebraic or perhaps topological group distinctions of their corresponding full groups (it is understood here that full groups are equipped with the uniform topology). For example, 3.17 provides such a distinction between hyperfinite and non-hyperfinite ones. One might explore the following possibility. Recall that an action of Γ on (X, µ) is called a free action if ∀γ 6= 1(γ · x 6= x, µ-a.e.). We know by Gaboriau [Ga1] that if Em is given by a free, measure preserving action of
26
I. MEASURE PRESERVING AUTOMORPHISMS
the free group Fm and En by a similar action of the free group Fn (1 ≤ m, n ≤ ∞), then if m 6= n, Em ∼ 6 En . Can one detect this by looking = at their full groups [Em ], [En ]? One can look for example at topological generators. Recall that given a topological group G a subset of G is a topological generator if it generates a dense subgroup of G. Denote by t(G) the smallest cardinality of a set of topological generators of G. (Thus t(G) = 1 iff G is monothetic.) Clearly 1 ≤ t(G) ≤ ℵ0 , if G is separable. In the notation above, and keeping in mind that if Γ is dense in ([E], u), then E = EΓX , Gaboriau’s [Ga1] theory of costs immediately implies that t([En ]) ≥ n. In fact, Ben Miller pointed out that one actually has t([En ]) ≥ n + 1. To see this, assume, towards a contradiction, that T1 , . . . , Tn are topological generators for [En ] and let Γ be the group they generate. Since En = EΓX , and En has cost n, it follows from Gaboriau [Ga1], I.11, that Γ acts freely. Let then S1 , S2 be two distinct elements of Γ and (as in the proof of 3.10) find disjoint Borel sets A1 , A2 of positive measure such that S1 (A1 ), S2 (A2 ) are also disjoint. Then there is S ∈ [En ] with S|Ai = Si |Ai , for i = 1, 2. Clearly, S cannot be in the uniform closure of Γ. In particular, t([E∞ ]) = ℵ0 . Note that, by 3.9, for any ergodic E we have t([E]) ≥ 2. In an earlier version of this work, I have raised the question of whether t([En ]) < ∞ (even in the case n = 1). This has now been answered by the following result. Theorem 4.12 (Kittrell-Tsankov [KiT]). An ergodic equivalence relation E is generated by an action of a finitely generated group iff t([E]) < ∞ (where [E] is equipped with the uniform topology). Moreover if En is generated by a free, measure preserving, ergodic action of Fn , then t([En ]) ≤ 3(n + 1). Kittrell and Tsankov [KiT] also proved that if n = 1, i.e., for ergodic hyperfinite E, we have t([E]) ≤ 3. It is unknown whether in this case the value of t([E]) is 2. For any n < ∞, we have n + 1 ≤ t([En ]) ≤ 3(n + 1). It is unknown if t([En ]) is independent of the action and, in case this has a positive answer, what is the exact value of t([En ]). In any case, the preceding shows that t([En ]) < t([Em ]), provided m + 1 > 3(n + 1). This appears to be the first result providing a topological group distinction between [Em ], [En ], provided m, n are sufficiently far apart. We remark that it is known that t(Aut(X, µ), w) = 2 and in fact the set of pairs (g, h) ∈ Aut(X, µ)2 that generate a dense subgroup is dense Gδ in Aut(X, µ)2 (see Grzaslewicz [Gr], Prasad [Pr], and Kechris-Rosendal [KR] for a different approach). Also it is not hard to see that if U (H) is the unitary group of an infinite dimensional Hilbert space, then t(U (H)) = 2, and again, in fact, the set of pairs (g, h) ∈ U (H)2 that generate a dense subgroup of U (H) is dense Gδ in U (H)2 . This is because if U (n) is the unitary group of the finite-dimensional Hilbert space Cn then (with some S canonical identifications) U (1) ⊆ U (2) ⊆ · · · ⊆ U (H) and n U (n) = U (H). Then since each U (n) is compact and connected, it follows, e.g., by a result
4. THE RECONSTRUCTION THEOREM
27
of Schreier-Ulam [SU], that the set of pairs (g, h) ∈ U (n)2 that generate a dense subgroup of U (n) is dense Gδ in U (n)2 . The same result S for U (H) follows, using the Baire Category Theorem, since for each u ∈ n U (n) and open nbhd N of u, the set of (g, h) ∈ U (H)2 that generate a subgroup intersecting N is open dense. It is perhaps worth pointing out here that although, as we have seen above, ([E], u) is topologically finitely generated, when E is given by an ergodic action of a finitely generated group, for any E and for any n ≥ 1 it is not the case that the set of n-tuples (T1 , . . . , Tn ) ∈ [E]n that generate a dense subgroup of ([E], u) is dense in [E]n (with the product uniform topology). To see this, simply notice that if T1 , . . . , Tn satisfy du (Ti , 1) < for i = 1, . . . , n, then there is a set A ⊆ X with µ(A) > 1 − n such that Ti |A = id, ∀i ≤ n, so T |A = id for each T ∈ hT1 , . . . , Tn i, thus if n < 1, hT1 , . . . , Tn i ⊆ {T ∈ [E] : du (T, 1) < n} is not dense in ([E], u). This simple argument also limits the kinds of Polish groups that can be closed subgroups of ([E], u) for an equivalence relation E. Call a Polish group G locally topologically finitely generated if there is n ≥ 1 such that for any open nbhd V of 1 ∈ G, there are g1 , . . . , gn ∈ G with hg1 , . . . , gn i dense in G. Examples of such groups include Rn , Tn , U (H), Aut(X, µ). Also factors and finite products of locally topologically finitely generated groups have the same property. Now notice that by the above argument any continuous homomorphism of such a group into the full group ([E], u) must be trivial. In particular, a non-trivial locally topologically finitely generated group cannot be a closed subgroup of ([E], u) (or even continuously embed into ([E], u)). (E) Recently Pestov (private communication) raised the following related question: Let E be a measure preserving, ergodic, hyperfinite equivalence relation. What kind of countable groups embed (algebraically) into [E]? In response to this we mention the following two facts. For the first, recall that a countable group Γ is residually finite (resp., residually amenable) if for any γ ∈ Γ, γ 6= 1 there is an homomorphism π : Γ → ∆, where ∆ is finite (resp., amenable) such that γ 6∈ ker(π). The following simple result was originally proved for residually finite groups. Ben Miller then noticed that the argument really shows the following stronger fact. Proposition 4.13. Let E be an ergodic, hyperfinite equivalence relation. Given a countable group Γ, if for every γ ∈ Γ \ {1} there is a homomorphism π : Γ → [E] such that γ 6∈ ker(π), then Γ embeds into [E]. In particular every residually amenable Γ embeds into [E] and for every countable group Γ there is unique normal subgroup N such that Γ/N embeds into [E] and every homomorphism from N into [E] is trivial. Proof. Suppose (X, µ) be the space on which E lives. Let Γ \ {1} = {γ1 , γ2 , . . . }. Fix then a sequence of homomorphisms πn : Γ → [E] such that γn 6∈ ker(πn ). Next split X into countably many Borel sets A1 , A2 , . . . , each meeting every E-class. Let En = E|An . Considering the space An with the normalized restriction of µ to An , En is hyperfinite and ergodic, so, by
28
I. MEASURE PRESERVING AUTOMORPHISMS
3.13, it is isomorphic to E, and thus, using πn , there is a Borel measure preserving action an of Γ on An such that γn acts non-trivially and the equivalence relation induced by an is contained in En . Taking the union of these actions on the An ’s gives an action on Γ on X such that every γ 6= 1 acts non-trivially and the equivalence relation induced by this action is included in E. This clearly gives an embedding of Γ into [E]. Since every amenable group can be clearly embedded into [E], by the result of Ornstein-Weiss [OW] and 3.13, it is clear that every residually amenable Γ embeds into [E]. Finally, the last assertion of the proposition is clear by taking N to be the intersection of the kernels of all homomorphisms from Γ into [E]. 2 We also have the following partial converse. It is a special case of a result of Robertson [Ro], concerning groups embeddable into the unitary group of the hyperfinite II1 factor but we give below a direct ergodic theory argument. Below we recall that a Borel (not necessarily countable) equivalence relation R on a standard Borel space Y is smooth if there is Borel map f : Y → Z, Z a standard Borel space, such that xRy ⇔ f (x) = f (y). If R is countable, this is equivalent to the existence of a Borel selector for R. Finally, if E is countable, measure preserving on (X, µ), then E is called smooth if its restriction to a co-null E-invariant Borel set is smooth in the previous sense. It is then easy to see that E is smooth iff E is finite. Proposition 4.14. Let E be an ergodic, hyperfinite equivalence relation. If Γ is a countable group, Γ ≤ [E] and Γ has property (T), then Γ is residually finite. Proof. Let F ⊆ E be the equivalence relation induced by Γ. Claim. F is smooth. Granting this, we can complete the proof as follows. Since F is smooth and measure preserving it must have finite classes a.e. Decompose then X into countably many Γ-invariant Borel sets A1 , A2 , . . . such that the Γorbits in An have cardinality n. Fixing a Borel linear ordering on X, we can identify each Γ-orbit in An with {1, 2, . . . , n} and thus each Γ-orbit in An gives rise to a homomorphism of Γ into Sn . Let γ ∈ Γ be different from 1. Then for some n, γ acts non-trivially on An , so the homomorphism corresponding to some Γ-orbit in An sends γ to something different from 1 in Sn . So Γ is residually finite. Proof of the claim. We will use the following result of SchmidtZimmer (see Zimmer [Zi1], 9.1.1): If a property (T) group Γ acts in an ergodic, measure preserving way on a standard measure space (Y, ν) and α : Γ × Y → Z is a Borel cocycle of this action, then α is a coboundary, i.e., α(γ, x) = f (γ · x) − f (x), for some Borel f : X → Z. Fix a free Borel action (n, x) 7→ n · x of Z on (X, µ) which generates E. To show that F is smooth, it is enough to show that if (Y, ν) is one of
4. THE RECONSTRUCTION THEOREM
29
the ergodic components of the action of Γ, with ν non-atomic, then F |Y is smooth. Consider the cocycle α from Γ × Y into Z : α(γ, x) = (the unique n such that γ · x = n · x). Then there is Borel f : Y → Z such that α(γ, x) = f (γ · x) − f (x). Fix a set A of positive ν-measure, thus meeting every F -class in Y , on which f is constant. Then if x, γ · x ∈ A we have α(γ, x) = 0, so γ · x = x, i.e., A meets every F -class in exactly one point, so F |Y is smooth. 2 There are related questions that came up in a discussion with Sorin Popa. What countable groups Γ embed (algebraically) into [E], for an ergodic, hyperfinite E, so that they generate E? More generally, what countable groups Γ have measure preserving, ergodic actions that generate a hyperfinite equivalence relation? Every group that has this last property does not have property (T) and one may wonder if conversely every non-property (T) group has such an action. This however fails. Fix a countable group Γ which is simple, not residually finite, and has property (T) (such groups exist as pointed out by Simon Thomas). Let F be a non-trivial finite group and let ∆ = Γ ∗ F . Then ∆ does not have property (T). Assume, towards a contradiction, that ∆ has an ergodic action which generates a hyperfinite equivalence relation E. Then there is a homomorphism π : ∆ → [E]. Clearly π is 1-1 or trivial on Γ. In the second case, π(∆) = π(F ) is a finite group generating E, which is absurd. In the first case Γ embeds into [E], which contradicts 4.14. On the other hand, one can see that the free groups embed into [E], E ergodic, hyperfinite, so that they generate E. Take, for example, F2 . It is enough to find S, T ∈ [E], which generate a free group that acts ergodically. But 3.9 shows that in the uniform topology the set of pairs (S, T ) in (APER∩ [E])2 that generate a free group is dense Gδ and, using 3.6, it is easy to check that the set of such pairs that generate a group acting ergodically is also dense Gδ . Thus the generic pair in (APER ∩ [E])2 works. In general, it is not clear what countable groups embed into the full group of a given ergodic equivalence relation, not necessarily hyperfinite. A result of Ozawa [O] implies that there is no full group [F ] into which every countable group embeds (algebraically). In particular, there is no full group [F ] that contains up to conjugacy all full groups or, equivalently, there is no equivalence relation F such that for any other equivalence relation E there is E 0 ⊆ F with E ∼ = E 0 . We give an ergodic theory proof of this fact in Section 14, (C). Ben Miller has pointed out that in the purely Borel theoretic context one actually has the opposite answer. Below all spaces are uncountable standard Borel spaces. If E, F are countable Borel equivalence relations on X, Y, E ∼ =B F means that there is a Borel bijection π : X → Y with xEy ⇔ π(x)F π(y). Also E vB F means that there is a Borel injection with xEy ⇔ π(x)F π(y). Miller shows that there is a countable Borel equivalence
30
I. MEASURE PRESERVING AUTOMORPHISMS
F such that for every countable Borel equivalence relation E there is E 0 ⊆ F with E ∼ =B E 0 . His argument goes as follows. Let E∞ be a countable Borel equivalence relation on X such that for any countable Borel equivalence relation E we have E vB E∞ (see DoughertyJackson-Kechris [DJK], 1.8). Let I(N) = N × N and put F = E∞ × I(N), an equivalence relation on Y = X × N. We will show that F works. Take first a non-smooth E (on some space Z). Let π : Z → X be Borel and 1-1 that witnesses E vB E∞ . Put π 0 (z) = (π(z), 0). Then π 0 witnesses E vB F . Let P = π 0 (Z). Clearly F |(Y \ P ) is compressible, via (x, n) 7→ (x, n + 1), so there is an aperiodic smooth subequivalence relation R ⊆ F |(Y \ P ) (see [DJK], 2.5). Let E 0 = (F |P ) ∪ R. Clearly E 0 ⊆ F , so it is enough to show that E 0 ∼ =B E ⊕ R and =B E. Now E 0 ∼ clearly there is an E-invariant Borel set A ⊆ Z such that E|A ∼ =B R⊕R⊕. . . (we are using here that E is not smooth). So E0 ∼ =B E ⊕ R ∼ =B (E|(Z \ A) ⊕ R ⊕ . . . ) ⊕ R ∼ =B (E|(Z \ A)) ⊕ (R ⊕ R ⊕ . . . ) ∼ =B E. Finally, when E is smooth, the proof is easy as F contains an aperiodic smooth subequivalence relation. Comments. The method of proof and most of the results about [E] up to 4.11 come (with some modifications) from the papers Eigen [Ei1, Ei2] and Fathi [Fa]. The numbers 5 and 10 in 4.2 and 4.5 can be replaced by 1, 3 resp., see Miller [Mi]. 5. Turbulence of conjugacy (A) Foreman and Weiss [FW] have shown that the conjugacy action of the group (Aut(X, µ), w) on (ERG, w) is turbulent. We will provide below (see 5.3) a different proof of (a somewhat stronger version of) this result. Our argument also shows that the conjugacy action of ([E], u) on (APER∩[E], u) is turbulent for any ergodic but not E0 -ergodic E. We will first give the proof of this for hyperfinite E in order to explain in a somewhat simpler context the main idea. In this section, we use Hjorth [Hj2] and Kechris [Kec3] as references for the basic concepts and results of Hjorth’s theory of turbulence. Theorem 5.1 (A special case of 5.2). Let E be a hyperfinite ergodic equivalence relation. Then the conjugacy action of ([E], u) on (APER ∩ [E], u) is turbulent. Proof. We have already seen in 3.4 that if T ∈ APER ∩ [E], then its conjugacy class (in [E]) is dense in (APER∩[E], u). We next verify that if T ∈ APER ∩ [E], then its conjugacy class (in [E]) is meager in (APER ∩ [E], u). First, by 2.5, the conjugacy class of T in (Aut(X, µ), w) is meager, so it is disjoint from a conjugacy-invariant dense Gδ in (Aut(X, µ), w), say C. We can also assume that C ⊆ APER by 2.3. By 3.14, D = C ∩ [E] 6= ∅.
5. TURBULENCE OF CONJUGACY
31
Clearly D ⊆ APER ∩ [E] is conjugacy invariant in [E] and, being nonempty, it is dense in (APER ∩ [E], u) by 3.4. Finally as w ⊆ u, C is Gδ in (Aut(X, µ), u) and so D is Gδ in (APER ∩ [E], u). As D is disjoint from the conjugacy class of T in [E], we are done. We finally verify that if T ∈ APER ∩ [E], then T is turbulent (for the conjugacy action of ([E], u) on (APER ∩ [E], u)). Fix > 0 and let U = {S ∈ APER ∩ [E] : δu0 (S, T ) < }. Fix also a nbhd V of 1 in ([E], u). It is enough to show that the local orbit O(T, U , V ) is dense in U . Since {gT g −1 : g ∈ [E]} is dense in (APER ∩ [E], u), it is clearly dense in (U , u). So fix any g ∈ [E] with gT g −1 ∈ U . It is enough to show that u gT g −1 ∈ O(T,SU , V ) . Let E = S ∞ n=1 En , where E1 ⊆ E2 ⊆ . . . are Borel with card([x]En ) ≤ n, ∀x. Then n [En ] is dense in ([E], u). Indeed, given S ∈ [E], if Xn = {x : S S(x) ∈ [x]En }, then X1 ⊆ X2 ⊆ . . . and n Xn = X. So for any ρ > 0, choose n large enough so that µ(Xn ) > 1 − ρ. Then, as x ∈ Xn ⇒ S(x) ∈ [x]En , we can find S 0 ∈ [En ] such that S|Xn = S 0 |Xn and so δu (S, S 0 ) < ρ. So we can clearly assume that g ∈ [En ], for some large enough n. Notice that it is enough to find a continuous path λ 7→ gλ in ([E], u), λ ∈ [0, 1], such that g0 = 1, g1 = g and gλ T gλ−1 ∈ U , ∀λ. Because then we ∈ V, ∀i < k. If can find λ0 = 0 < λ1 < · · · < λk = 1 with gλi+1 gλ−1 i −1 −1 T1 = gλ1 , T2 = gλ2 gλ1 , . . . , Tk = gλk gλk−1 , then Ti ∈ V and −1 −1 ∈ U , ∀i ≤ k. Ti = gλi T gλ−1 Ti Ti−1 . . . T1 T T1−1 . . . Ti−1 i
Choose δ0 small enough so that if α = δu0 (gT g −1 , T ), then α + 4δ0 < . Find then N ≥ n large enough, so that there is S ∈ [EN ] with δu0 (S, T ) < δ0 . Then δu0 (S, gSg −1 ) < δ0 + δu0 (T, gT g −1 ) + δ0 = α + 2δ0 . Claim. There is a continuous path λ 7→ gλ in ([E], u) with g0 = 1, g1 = g and δu0 (S, gλ Sgλ−1 ) < α + 2δ0 . Then δu0 (T, gλ T gλ−1 ) < δ0 + α + 2δ0 + δ0 = α + 4δ0 < , i.e., gλ T gλ−1 ∈ U , ∀λ. Proof of the claim. Let Z be a Borel transversal for EN and let λ 7→ Zλ be continuous from [0, 1] into MALGµ , so that Z0 = ∅ ⊆ Zλ ⊆ Zν ⊆ Z1 = Z, µ(Zν \ Zλ ) ≤ (ν − λ), for 0 ≤ λ ≤ ν ≤ 1. Let Xλ = [Zλ ]EN , so that µ(Xν \ Xλ ) ≤ N · (ν − λ). Finally, define ( g on Xλ , gλ = id on X \ Xλ .
32
I. MEASURE PRESERVING AUTOMORPHISMS
Then λ 7→ gλ is a continuous path in ([E], u) and g0 = 1, g1 = g. Moreover, ¯ λ = id|X ¯ λ , where X ¯ λ = X \ Xλ . Xλ is gλ , S-invariant for each λ and gλ |X −1 0 We check that δu (S, gλ Sgλ ) < α + 2δ0 . Fix A ∈ MALGµ . Then ¯ λ )]∆ gλ Sgλ−1 (A)∆S(A) = [gλ Sgλ−1 (A ∩ Xλ ) ∪ gλ Sgλ−1 (A ∩ X ¯ [S(A ∩ Xλ ) ∪ S(A ∩ Xλ )] ¯ λ )]∆ = [gλ Sgλ−1 (A ∩ Xλ ) ∪ S(A ∩ X ¯ λ )] [S(A ∩ Xλ ) ∪ S(A ∩ X −1 ⊆ gλ Sgλ (A ∩ Xλ )∆S(A ∩ Xλ ) = [gSg −1 (A)∆S(A)] ∩ Xλ , so µ(gλ Sgλ−1 (A)∆S(A)) ≤ µ(gSg −1 (A)∆S(A)) ≤ δu0 (gSg −1 , S) < α + 2δ0 and we are done. 2 Remark. I would like to thank Greg Hjorth for suggesting a modification that simplified considerably my original calculation in the preceding proof. We next show that a variation of the proof of 5.1 allows one to show the following stronger statement. Recall that an equivalence relation E is E0 -ergodic (or strongly ergodic) if for every Borel homomorphism π : E → E0 , (i.e., a Borel map π : X → 2N with xEy ⇒ π(x)E0 π(y)) the preimage of some E0 -class is conull. If E = EΓX for a Borel action of a countable group Γ on X and E is E0 -ergodic, we will also call this action E0 -ergodic. See Jones-Schmidt [JS] and Hjorth-Kechris [HK3], Appendix A, for some basic properties of this concept. Theorem 5.2 (Kechris). Let E be an ergodic equivalence relation which is not E0 -ergodic. Then the conjugacy action of ([E], u) on (APER ∩ [E], u) is turbulent. Proof. Since E is not E0 -ergodic, there is a Borel homomorphism π : E → S∞ E0 such that the preimage of every E0 -class is µ-null. Write E0 = n=1 S En , E1 ⊆ E2 ⊆ . . . , En Borel with finite classes. We claim that the set n [π −1 (En ) ∩ E] is dense in ([E], u). Indeed, if PER denotes the set of periodic elements of Aut(X, µ), i.e., S ∈ PER ⇔ ∀x∃n(T n (x) = x), then, by 3.3, PER∩[E] is dense in ([E], u). So fix a periodic S ∈ [E] and let Xn = {x : π([x]S ) is contained in a single En -class}, S where [x]S is the S-orbit of x. Then X1 ⊆ X2 ⊆ . . . and X = n Xn . Also each Xn is S-invariant and if Sn = S|Xn ∪ id|(X \ Xn ), then Sn ∈ [π −1 (En ) ∩ E] and Sn → S uniformly. As in the proof of 5.1, whose notation we keep below, fix T ∈ APER ∩ [E], U , and g ∈ [π −1 (En ) ∩ E] with gT g −1 ∈ U . Then choose δ0 small enough so that α + 4δ0 < and N ≥ n large enough so that for some S ∈ [π −1 (EN ) ∩ E], δu0 (S, T ) < δ0 . It is again enough to find a continuous path λ 7→ gλ with g0 = 1, g1 = g and δu0 (S, gλ Sgλ−1 ) < α + 2δ0 .
5. TURBULENCE OF CONJUGACY
33
Fix a Borel transversal A for EN and let ν = π∗ µ. Then ν(C) = 0 for each EN -class C. Define the measure ρ on A by ρ(B) = ν([B]EN ), B ⊆ A Borel. Clearly it is non-atomic, so there is a continuous map λ 7→ Aλ from [0, 1] to MALGρ with A0 = ∅, A1 = A and 0 ≤ λ ≤ λ0 ≤ 1 ⇒ Aλ ⊆ Aλ0 and ρ(Aλ0 \ Aλ ) ≤ λ0 − λ. Let Xλ = π −1 ([Aλ ]EN ), so that µ(Xλ ) = ρ(Aλ ), thus λ 7→ Xλ is continuous from [0, 1] to MALGµ , X0 = ∅, X1 = X, 0 ≤ λ ≤ λ0 ≤ 1 ⇒ Xλ ⊆ Xλ0 and µ(Xλ0 \Xλ ) ≤ λ0 −λ. Moreover Xλ is π −1 (EN )-invariant, so g, S-invariant. Then define as before gλ = g|Xλ ∪ id|(X \ Xλ ), so that Xλ is also gλ , S-invariant. The proof then proceeds exactly as in 5.1. 2 With essentially the same proof one can show the result of ForemanWeiss that the conjugacy action of (Aut(X, µ), w) on (ERG, w) is turbulent. Actually one can prove a somewhat more precise version (replacing ERG by APER) by using an additional fact (Lemma 5.4 below). Theorem 5.3 (Foreman-Weiss [FW] for ERG). The conjugacy action of the group (Aut(X, µ), w) on (APER, w) is turbulent. Proof. We have already seen that every conjugacy class in APER is dense and meager. We finally fix T ∈ APER in order to show that it is turbulent for the conjugacy action. The following lemma was proved together with Ben Miller. Lemma 5.4. Every aperiodic, hyperfinite equivalence relation E is contained in an ergodic, hyperfinite equivalence relation F . Granting this, let S ∈ ERG be such that [T ] ⊆ [S], where for T ∈ Aut(X, µ) we denote by [T ] = [ET ] the full group of the equivalence relation ET induced by T . Now [S] is dense in (Aut(X, µ), w). Then repeat the argument in the proof of 5.1, starting with g ∈ [S]. Proof of the Lemma. Consider the ergodic decomposition of E. This is given by a Borel E-invariant surjection π : X → E, where E is the space of ergodic invariant measures for E. If Xe = {x ∈ X : π(x) = e}, then e is the unique invariant ergodic R measure for E|Xe and if ν = π∗ µ, a probability measure on E, then µ = edν(e), i.e., for A ⊆ X Borel, Z µ(A) = e(A ∩ Xe )dν(e). In particular, modulo µ-null sets, every E-invariant set A is of the form π −1 (B), for some B ⊆ E. Case 1. E is finite. We will prove then the result by induction on card(E). If card(E) = 1, there is nothing to prove. Assume now card(E) = 2, say E = {e1 , e2 }, with µ(Xe1 ) ≥ µ(Xe2 ). Note that ei = (µ|Xei )/µ(Xei ). By Dye’s Theorem, we can find X1 ⊆ Xe1 , with µ(X1 ) = µ(Xe2 ) and a Borel isomorphism ϕ from E|X1 to E|Xe2 (modulo null sets) which preserves µ. Then define F to
34
I. MEASURE PRESERVING AUTOMORPHISMS
be the equivalence relation generated by E and ϕ. Clearly F is measure preserving and ergodic. Since every F -class consists of at most 2 E-classes, F is hyperfinite. Assume now that the result is true up to n and say E = {e1 , . . . , en , en+1 }. By induction hypothesis we can find a hyperfinite measure preserving ergodic equivalence relation En on Xe1 ∪ · · · ∪ Xen extending E|(Xe1 ∪ · · · ∪ Xen ). Then if Fn = E ∪En , Fn is hyperfinite, measure preserving and has 2 ergodic invariant measures, so by the n = 2 case, we can find F ⊇ Fn , F hyperfinite, ergodic and measure preserving. Case 2. E is countably infinite. Say E = {e1 , e2 , . . . ). By Case 1, there is hyperfinite, measure preserving, ergodic ES n on Xe1 ∪· · ·∪Xen with E|Xe1 ∪· · ·∪E|Xen ⊆ En and En ⊆ En+1 . Let F = n En . This clearly works. Case 3. E is uncountable and ν is non-atomic. Let then T ∈ Aut(E, ν) be ergodic. For each e ∈ E, let ϕe : Xe → XT (e) be a Borel bijection such that ϕ(e, x) = ϕe (x) is Borel and ϕe is an isomorphism of (E|Xe , e) with (E|XT (e) , T (e)) modulo null sets. Let then ϕ(x) = ϕπ(x) (x) (so if x ∈ Xe , ϕ(x) = ϕe (x)). Note first that ϕ ∈ Aut(X, µ). To see this fix a Borel set A ⊆ X. Then Z −1 µ(ϕ (A)) = e(ϕ−1 (A) ∩ Xe )dν(e) Z = T (e)(ϕe (ϕ−1 (A) ∩ Xe ))dν(e) Z = T (e)(A ∩ XT (e) )dν(e) Z = e(A ∩ Xe )dν(e) = µ(A), as ν is T -invariant. Let F be the equivalence relation induced by E and ϕ, so that E ⊆ F and F is measure preserving, ergodic. Finally, E is hyperfinite (µ-a.e.) as every F class can be (Borel uniformly) ordered in order type ζ 2 (where ζ is the order type of Z) (see Jackson-Kechris-Louveau [JKL], Section 2). Case 4. E is uncountable but ν has atoms. S Let then {e1 , e2 , . . . } be the atoms of ν. Let X0 = i Xei , X1 = X \ X0 . By applying the preceding cases to (E|Xi , (µ|Xi )/µ(Xi )), i = 0, 1, we see that we can find measure preserving, ergodic, hyperfinite Fi on Xi such that E|Xi ⊆ Fi . If F 0 = F1 ∪ F2 , then F is measure preserving, hyperfinite and E ⊆ F 0 . Then, by Case 1 again, we can find measure preserving, ergodic, hyperfinite F ⊇ F 0 and the proof is complete. 2 Similar arguments also show the following.
5. TURBULENCE OF CONJUGACY
35
Theorem 5.5. Let E be an ergodic equivalence relation. Then the conjugacy action of ([E], u) on (Aut(X, µ), w) is generically turbulent. Proof. Use the argument in 5.1 to show that every ergodic T ∈ [E] is turbulent for this action (since the conjugacy class of T in Aut(X, µ) is weakly dense, this shows that the action is generically turbulent). 2 (B) Given equivalence relations E, F on standard Borel spaces X, Y , we say that E can be Borel reduced to F , in symbols E ≤B F, if there is a Borel map f : X → Y such that xEy ⇔ f (x)F f (y). We say that E is Borel bireducible to F , in symbols E ∼B F, if E ≤B F and F ≤B E. An equivalence relation E on a standard Borel space X can be classified by countable structures if there is a countable language L and a Borel map f : X → XL , where XL is the standard Borel space of countable structures for L (with universe N), such that xEy ⇔ f (x) ∼ = f (y), where ∼ = is the isomorphism relation for structures, i.e, E is Borel reducible to the isomorphism relation of the countable structures of some countable language. Hjorth [Hj2] has shown that if E is induced by a (generically) turbulent action, then E restricted to any dense Gδ set cannot be classified by countable structures. From 5.3 and the fact that WMIX is dense Gδ one can of course derive that conjugacy in WMIX cannot be classified by countable structures. Earlier such a result for ERG was proved in Hjorth [Hj1]. Moreover, by also using 2.5, it also follows that unitary equivalence in WMIX cannot be classified by countable structures. Theorem 5.6 (Hjorth [Hj1] for ERG, Foreman-Weiss [FW]). Conjugacy and unitary (spectral) equivalence in WMIX cannot be classified by countable structures. In fact one can prove stronger results which apply as well to MIX (which is meager in Aut(X, µ)). Fix an uncountable standard Borel space Y and let P (Y ) be the standard Borel space of Borel probability measures on Y . As usual call µ, ν ∈ P (Y ) equivalent measures (or mutually absolutely continuous) if they have the same null sets. Thus denoting equivalence of µ, ν by µ ∼ ν and absolute continuity by µ ν, we have µ ∼ ν ⇔ µ ν and ν µ. Clearly, up to Borel isomorphism, ∼ is independent of the choice of Y . It is known that ∼ is a Borel equivalence relation which cannot be classified by countable structures, see Kechris-Sofronidis [KS]. The spectral theorem for
36
I. MEASURE PRESERVING AUTOMORPHISMS
unitary operators implies that if U (H) is the unitary group of a separable infinite dimensional space H, then conjugacy in U (H) is Borel bireducible to ∼. It is clear then that unitary (spectral) equivalence in Aut(X, µ) is Borel reducible to measure equivalence ∼. Theorem 5.7 (Kechris). (a) Measure equivalence is Borel bireducible to unitary equivalence in MIX (and thus also in WMIX, ERG and Aut(X, µ)). In particular, unitary equivalence in MIX cannot be classified by countable structures. (b) Measure equivalence is Borel reducible to conjugacy in MIX (and thus also in WMIX, ERG and Aut(X, µ)). In particular, conjugacy in MIX cannot be classified by countable structures. Our proof will use the spectral theory of unitary operators and measure preserving transformations, for which some standard references are Cornfeld-Fomin-Sinai [CFS], Glasner [Gl2], Goodson [Goo], Lema´ nczyk [Le], Nadkarni [Na], Parry [Pa], Queff´elec [Qu], and in particular the spectral theory of the shift in Gaussian spaces (see Appendices C-E) for which we refer the reader to Cornfeld-Fomin-Sinai [CFS]. Moreover, we will also make use of some results in the harmonic analysis of measures on T, for which we refer to Kechris-Louveau [KL]. Proof of 5.7. (a) For each finite (positive) Borel measure σ on T consider σ ˆ : Z → C given by Z σ ˆ (n) = z¯n dσ(z). By Herglotz’s theorem (see, e.g., Parry [Pa]) σ ↔ σ ˆ is a 1-1 correspondence between finite Borel measures on T and positive-definite functions on Z. Note that σ ˆ is real iff σ is symmetric, i.e., invariant under conjugation ¯ for every Borel set A, where A¯ = {¯ (σ(A) = σ(A), z : z ∈ A}). To each symmetric σ then we can associate the real positive-definite function ϕ(σ) = σ ˆ and then the Gaussian space (RZ , µϕ(σ) ), as in Appendix C, and the corresponding shift, which we will denote by Tσ , i.e., Tσ ((xn )n∈Z ) = (xn−1 )n∈Z . It is well-known (see, Cornfeld-Fomin-Sinai [CFS], p. 369) that Tσ is mixing iff σ is a Rajchman measure, i.e., σ ˆ (n) → 0 as |n| → ∞. Consider the Wiener chaos decomposition M L2 (RZ , µϕ(σ) ) = L2 (RZ , µϕ(σ) , C) = HC:n: n≥0
discussed in Appendix D, and let UTσ = Uσ be the unitary operator on L2 (RZ , µϕ(σ) ) induced by Tσ . Clearly each HC:n: is invariant under Uσ . Also note that HC:0: = C is the subspace of constants and M L20 (RZ , µϕ(σ) ) = HC:n: . n≥1
5. TURBULENCE OF CONJUGACY
37
For each finite measure σ on T consider the Hilbert space L2 (T, σ) and the unitary operator Vσ (f )(z) = zf (z). Suppose now that Λ ⊆ T is an independent set, i.e., if z1 , . . . , zk ∈ Λ are distinct, n1 , . . . , nk ∈ Z, and z1n1 · · · zknk = 1, then n1 = · · · = nk = 0. If a symmetric σ is supported ¯ then the n-fold convolution σ ∗n = σ ∗ · · · ∗ σ is supported by by Λ ∪ Λ, n ¯ (= {z1 . . . zn : zi ∈ (Λ ∪ Λ)}). ¯ (Λ ∪ Λ) Moreover if σ is non-atomic, then ∗n ∗n ¯ ¯ zi 6= actually σ is supported by (Λ ∪ Λ) = {z1 . . . zn : zi ∈ (Λ ∪ Λ), ∗1 ∗2 zk , z¯k , if i 6= k} and these sets are pairwise disjoint, so σ = σ , σ , . . . have pairwise disjoint supports. Now the spectral theory of Gaussian shifts shows that the unitary operator Uσ on HC:n: , n ≥ 1, is isomorphic to the unitary operator Vσ∗n on L2 (T, σ ∗n ); see Cornfeld-Fomin-Sinai [CFS], P Chapter 14, 1 ∗n §4, Theorem 1. If σ is also a probability measure, let σ∞ = ∞ n=1 2n σ . L 2 2 Then L (T, σ∞ ) = n≥1 Kn , where Kn = {f ∈ L (T, σ∞ ) : f is 0 off Λ∗n }, each Kn is invariant under Vσ∞ and the map f ∈ Kn 7→ √12n f ∈ L2 (T, σ ∗n ) is an isomorphism of Vσ∞ |Kn with Vσ∗n . It follows that Uσ |L20 (RZ , µϕ(σ) ) is isomorphic to Vσ∞ on L2 (T, σ∞ ). In the language of spectral theory this says that the maximal spectral type of the operator Uσ0 = Uσ |L20 (RZ , µϕ(σ) ) is equal to σ∞ and it has multiplicity 1. Since, by the spectral theory, two unitary operators are isomorphic iff their maximal spectral types are equivalent and they have the same multiplicity function, we conclude that whenever σ, τ are non-atomic symmetric probability Borel measures on T ¯ Λ independent, and U 0 , U 0 are each supported by a set of the form Λ ∪ Λ, σ τ the unitary operators associated to the respective Gaussian shifts (restricted to the orthogonal of the constant functions), then σ∞ ∼ τ∞ ⇔ Uσ0 ∼ = Uτ0 , where ∼ = denotes isomorphism of the unitary operators. ¯ But one can now verify that if σ, τ are supported by the same set Λ ∪ Λ, where Λ is an independent set, then σ ∼ τ ⇔ σ ∞ ∼ τ∞ . Indeed, σ ∼ τ ⇒ σ∞ ∼ τ∞ holds simply because σ << τ ⇒ σ ∗ ρ << τ ∗ ρ, for any ρ. Conversely, assume that σ∞ ∼ τ∞ . Say σ, τ are both supported ¯ Suppose now σ(A) = 0 but τ (A) > 0, towards a contradiction. by Λ ∪ Λ. ¯ we can assume that A ⊆ (Λ ∪ Λ). ¯ Replacing A by A ∩ (Λ ∪ Λ), Since ∗n σ∞ ∼ τ∞ >> τ , we have σ∞ (A) > 0, so as σ(A) = 0, σ (A) > 0 for some ¯ ∗n 6= ∅ and therefore (Λ ∪ Λ) ¯ ∩ (Λ ∪ Λ) ¯ ∗n 6= ∅, a n > 1. Thus A ∩ (Λ ∪ Λ) ¯ contradiction. Thus if σ, τ are supported by Λ ∪ Λ for the same independent set Λ, then σ ∼ τ ⇔ Uσ0 ∼ = Uτ0 . Note also that Uσ0 ∼ = Uτ0 ⇔ Uσ ∼ = Uτ , so we finally have:
38
I. MEASURE PRESERVING AUTOMORPHISMS
If σ, τ are symmetric Rajchman (and thus non-atomic) probability mea¯ for the same independent set Λ, then Tσ , Tτ are sures supported by Λ ∪ Λ mixing and σ ∼ τ ⇔ Tσ , Tτ are unitarily equivalent. As σ 7→ Tσ is clearly Borel, we have shown that ∼ restricted to symmetric ¯ for a fixed independent Rajchman probability measures supported by Λ ∪ Λ, set Λ, is Borel reducible to unitary equivalence on MIX. It only remains to show that ∼ can be Borel reduced to ∼ restricted to such measures. We will use here the following result of Rudin [Ru1] (see also KahaneSalem [KS], VIII, 3, Th´eor`eme II): There is a closed set Λ ⊆ T which is independent and supports a Rajchman measure, i.e., Λ is in the class M0 of closed sets of restricted multiplicity (see Kechris-Louveau [KL]). By a theorem of Kaufman (see Kechris-Louveau [KL], VII. 1, Theorem 7), there is a Borel function x 7→ σx from 2N to the standard Borel space of probability measures on T such that each σx is a Rajchman measure, supp(σx ) ∩ supp(σy ) = ∅ if S x 6= y (where supp(σ) = T \ {U open : σ(U ) = 0}), and supp(σx ) ⊆ Λ. Then, since Λ is independent, we have A∩B = ∅, for disjoint A, B ⊆ Λ, so, by ¯ we have the following conreplacing σx by 12 (σx + σ ¯x ) (where σ ¯ (A) = σ(A)), clusion: There is a Borel map x 7→ σx from 2N to symmetric Rajchman prob¯ and supp(σx ) ∩ supp(σy ) = ∅ ability Borel measures with supp(σx ) ⊆ Λ ∪ Λ if x 6= y. Now consider an arbitrary probability Borel measure µ on 2N . Put Z σµ = σx dµ(x) R Ri.e., σµ (A)R = σx (A)dµ(x), for each Borel set A ⊆ T, or equivalently f dσµ = (f dσx )dµ(x) for every continuous function f on T. It follows that Z σ ˆµ (n) = σ ˆx (n)dµ(x) and since σ ˆx (n) = fn (x) → 0 as |n| → ∞ and |fn (x)| ≤ 1 we have, by Lebesgue Dominated Convergence, that σ ˆµ (n) → 0, i.e., σµ is a Rajchman ¯ It only measure. Clearly σµ is symmetric and σµ is supported by Λ ∪ Λ. remains to verify that µ ν ⇔ σµ σν . Assume µ ν and let A ⊆ T be Borel with σν (A) = 0, so that Z σx (A)dν(x) = 0. Then σx (A) = 0, ν−a.e.(x), thus σx (A) = 0, µ−a.e.(x), therefore σµ (A) = 0. Thus σµ σν . Conversely,Sassume that σµ σν and let X ⊆ 2N be Borel with ν(X)R= 0. Let X ∗ = x∈X supp(σx ), which is a Borel set in T. Then R σν (X ∗ ) = σx (X ∗ )dν(x) = ν(X) = 0. Thus σµ (X ∗ ) = σx (X ∗ )dµ(x) = 0, so σx (X ∗ ) = 0, µ−a.e.(x), therefore x 6∈ X, µ−a.e.(x), i.e., µ(X) = 0. So µ ν.
5. TURBULENCE OF CONJUGACY
39
This concludes the proof of (a). (b) By the proof in part (a), it is clearly enough to show that if σ, τ are symmetric measures on T, and σ ∼ τ , then Tσ , Tτ are isomorphic transformations, i.e., there is an isomorphism S : (RZ , µϕ(σ) ) → (RZ , µϕ(τ ) ) sending Tσ to Tτ . Suppose σ ∼ τ . Then the map f ∈ L2 (T, σ) 7→
dσ f ∈ L2 (T, τ ) dτ
is an isomorphism of the Hilbert space L2 (T, σ) to L2 (T, τ ) that sends Vσ to Vτ . Since σ, τ are both symmetric, i.e., invariant under conjugation, so is dσ dτ , dσ dσ 2 2 i.e., dτ (z) = dτ (¯ z ) (τ −a.e.). Let S (T, σ) be the closed subset of L (T, σ) consisting of all f ∈ L2 (T, σ) satisfying f (¯ z ) = f (z). This is a real subspace of L2 (T, σ) invariant under the operator Vσ and L2 (T, σ) is the complexification of S 2 (T, σ), since any f ∈ L2 (T, σ) is uniquely written as f1 + if2 , 1 with f1 , f2 ∈ S 2 (T, σ), namely f1 (z) = 21 [f (z) + f (¯ z )], f2 (z) = 2i [f (z) − 2 f (¯ z )]. Moreover, Vσ on L (T, σ) is the complexification of Vσ |S 2 (T, σ). Now the spectral theory of the Gaussian shift Tσ asserts that there is an isomorphism between the real Hilbert space Hσ = H :1: (the first chaos of L2 (RZ , µϕ(σ) , R) associated with the Gaussian space (RZ , µϕ(σ) )) and the real Hilbert space S 2 (T, σ), which sends Uσ |Hσ to Vσ |S 2 (T, σ) (see CornfeldFomin-Sinai [CFS], p. 368). Similarly for τ . Since dσ dτ is invariant under dσ 2 2 conjugation, the map f ∈ L (T, σ) 7→ dτ f ∈ L (T, τ ) sends S 2 (T, σ) to S 2 (T, τ ) and Vσ |S 2 (T, σ) to Vτ |S 2 (T, σ). Thus there is an isomorphism T between Hσ and Hτ that sends Uσ |Hσ to Uτ |Hτ . By Appendix D, there is an isomorphism S : (RZ , µϕ(σ) ) → (RZ , µϕ(τ ) ) such that if OS (f ) = f ◦ S −1 is the corresponding isomorphism of L2 (RZ , µϕ(σ) , R) with L2 (RZ , µϕ(τ ) , R), then OS |Hσ = T . Since T sends Uσ |Hσ to Uτ |Hτ , S sends Tσ to Tτ and the proof is complete. 2 (C) Very recently Foreman, Rudolph and Weiss [FRW] have shown that the conjugacy equivalence relation on ERG is not Borel (in fact it is a complete Σ11 subset of ERG2 ). This in particular shows that conjugacy on ERG cannot be Borel reduced to unitary equivalence on ERG and thus, in combination with 5.7, it shows that, in terms of Borel reducibility, conjugacy on ERG is strictly more complicated that unitary equivalence on ERG. Although the place of unitary equivalence on ERG in the Borel reducibility hierarchy of complexity of equivalence relations is completely understood, it is still open to determine this for conjugacy on ERG. For example, how does it compare with the universal equivalence relation induced by a Borel action of (Aut(X, µ), w) (or equivalently of U (H), which is isomorphic to a closed subgroup of (Aut(X, µ), w), by Appendix E)? Another question (motivated by a discussion with Yehuda Shalom) on which not much seems to be known is the following. It has been a classical
40
I. MEASURE PRESERVING AUTOMORPHISMS
problem of ergodic theory to distinguish up to conjugacy ergodic transformations which are unitarily equivalent. The concept of entropy provided a powerful tool for attacking this problem. In this vein one can raise the general problem of understanding the complexity of conjugacy within each unitary equivalence class. More precisely, let C ⊆ ERG denote a given unitary equivalence class in ERG. How complicated is conjugacy restricted to C? The answer will depend of course on C. If C corresponds to a discrete spectrum ergodic transformation T (i.e., one for which UT has discrete spectrum, which means that the corresponding maximal spectral type concentrates on a countable set), then, by the classical Halmos-von Neumann Theorem, C consists of a single conjugacy class. On the other hand not much seems to be known about the complexity of conjugacy for C whose maximal spectral type is continuous (i.e., non-atomic). A particular case of interest is the class C which has countable homogeneous Lebesgue spectrum (i.e., the unitary equivalence class of the shift on 2Z with the usual product measure), which of course contains continuum many conjugacy classes. Remark. Concerning the spectral theory of ergodic measure preserving transformations, it is a very hard and still unsolved problem to determine what are the possible spectral invariants, i.e., pairs of measure classes in P (T) and multiplicity functions, that correspond to (the unitary operator associated to) an ergodic measure preserving transformation. Equivalently, this can be viewed as the question of characterizing the unitary operators that are realized by ergodic, measure preserving transformations up to isomorphism (i.e., conjugacy in the unitary group) - see here also Appendix H, (F). Fox example, one can ask if the set of such of operators is Borel in U (H). (It is obviously Σ11 .) It also appears to be unknown what measure classes of probability measures on T can appear as maximal spectral types of ergodic measure preserving transformations. (See, e.g., Katok-Thouvenot [KTh].) For example, one can ask whether the set of maximal spectral types of such transformations is a Borel set in P (T). (It is clearly Σ11 .) 6. Automorphism groups of equivalence relations (A) For each measure preserving countable Borel equivalence relation E on (X, µ) we denote by N [E] the group of all T ∈ Aut(X, µ) such that xEy ⇔ T (x)ET (y), for all x, y in a conull set. Note that N [E] is the normalizer of [E] in Aut(X, µ). If E is not smooth, then T.-J. Wei [We] showed that N [E] is Π03 -complete in (Aut(X, µ), w). Clearly N [E] is closed in (Aut(X, µ), u) but it may not be separable. For example, if E = E0 and for A ⊆ N we let fA (x) = x + χA , where addition is pointwise modulo 2, then fA ∈ N [E] and δu (fA , fB ) = 1 if A 6= B. Next we will see that N [E] is a Polishable subgroup of (Aut(X, µ), w), if E is aperiodic.
6. AUTOMORPHISM GROUPS OF EQUIVALENCE RELATIONS
41
Note that each T ∈ N [E] induces by conjugation an isometry iT of ([E], δu ) : iT (S) = T ST −1 . Consider the Polish group Iso([E], δu ), with the pointwise convergence topology and the map i(T ) = iT . We first check that it is an algebraic isomorphism of N [E] with a subgroup of Iso([E], δu ), provided that E is aperiodic. It is clearly a homomorphism. Injectivity follows from 4.11. We next verify that the image i(N [E]) of N [E] is closed in Iso([E], δu ), provided E is aperiodic. To see this let Tn ∈ N [E] and assume iTn → i0 in Iso([E], δu ). Then for every S ∈ [E], {Tn STn−1 } is δu -Cauchy, i.e., −1 −1 lim δu ((Tm Tn )S(Tm Tn )−1 , S) → 0.
m,n→∞
Fix A ∈ MALGµ and find S ∈ [E] with supp(S) = A (by 4.10). Since −1 T )S(T −1 T )−1 ) = T −1 T (A), we have µ(T −1 T (A)∆A) → 0, supp((Tm n m n m n m n as m, n → ∞. Similarly, since also (iTn )−1 = iTn−1 → i−1 0 in Iso([E], δu ), we have limm,n→0 µ(Tm Tn−1 (A)∆A) = 0. So limm,n→∞ [µ(Tn (A)∆Tm (A)) + −1 (A)] → 0, i.e., {T } is a Cauchy sequence in the complete µ(Tn−1 (A)∆Tm n ¯ metric δw of Aut(X, µ). Thus there is T ∈ Aut(X, µ) with Tn → T weakly. Then for S ∈ [E], Tn STn−1 → T ST −1 weakly. But also Tn STn−1 → i0 (S) uniformly, so T ST −1 = i0 (S) ∈ [E]. Thus T ∈ N [E] and clearly iT = i0 . We summarize the preceding discussion as follows. Theorem 6.1. Let E be aperiodic. The group N [E] is a Polishable subgroup of (Aut(X, µ), w). The corresponding Polish topology τN [E] is given by the complete metric X δN [E] (T1 , T2 ) = 2−n [δu (T1 Sn T1−1 , T2 Sn T2−1 ) + δu (T1−1 Sn T1 , T2−1 Sn T2 )], n
where {Sn } is dense in ([E], u). We have w|N [E] ⊆ τN [E] ⊆ u|N [E], and thus if Tn → T in N [E], then Tn → T weakly. Convention. From now on, when we consider topological properties of N [E] without explicitly indicating the topology, we will always assume that we refer to τN [E] . If E = EΓX , where the countable group Γ acts in a Borel way on X, we also have the following fact, denoting for each γ ∈ Γ also by γ the map γ(x) = γ · x. Proposition 6.2. For E = EΓX aperiodic, Tn , T ∈ N [E], Tn → T ⇔ Tn → T weakly and ∀γ ∈ Γ(Tn γTn−1 → T γT −1 uniformly). Proof. Again it is enough to show that if Tn ∈ N [E], Tn → 1 weakly and Tn γTn−1 → γ uniformly, ∀γ ∈ Γ, then Tn T Tn−1 →ST uniformly, ∀T ∈ [E]. Fix T ∈ [E]. Then there is a partition X = i Ai , Ai Borel, and γi ∈ Γ with T (x) = γi (x), ∀x ∈ Ai . Fix > 0. Choose then M large enough so that
42
I. MEASURE PRESERVING AUTOMORPHISMS
P
i>M
µ(Ai ) < . Then for some N and all n ≥ N , µ(Tn (Ai )∆Ai ) < , ∀i ≤ M, (M + 1)
Tn γi Tn−1 (x) = γi (x), ∀i ≤ M, for all x 6∈ Ain , where µ(Ain ) < (M+1) , i ≤ M . Thus off a set of measure < 2, we have that if i ≤ M and x ∈ Ai , then Tn−1 (x) ∈ Ai and T Tn−1 (x) = γi (Tn−1 (x)) = Tn−1 (γi (x)) = Tn−1 T (x). Thus µ({x : T Tn−1 (x) 6= Tn−1 T (x)}) < 3, if n ≥ N , so δu (T Tn−1 , Tn−1 T ) = δu (Tn T Tn−1 , T ) → 0. 2 Thus the following is also a complete metric for N [E]: X δ¯N [E] (T1 , T2 ) = δ¯w (T1 , T2 ) + 2−n δu (T1 γn T1−1 , T2 γn T2−1 ), n
where Γ = {γn }. Remark. It is clear that in the preceding one can equivalently use δu0 instead of δu . (B) There is also another way to understand the topology of N [E]. Consider the measure M on E defined by Z M (A) = card(Ax )dµ(x), where Ax = {y : (x, y) ∈ A}, for any Borel set A ⊆ E. This is a σ-finite Borel measure on E. Given now T ∈ N [E] consider the map T × T on E defined by T × T (x, y) = (T (x), T (y)). Then it is easy to see that T × T preserves the measure M , therefore induces a unitary operator on the Hilbert space L2 (E, M ) denoted by UT ×T : UT ×T (f )(x, y) = f (T −1 (x), T −1 (y)). The map T 7→ UT ×T is a group isomorphism between N [E] and a subgroup of U (L2 (E, M )). (To see that it is 1-1, assume that UT ×T = 1, i.e., f (T −1 (x), T −1 (y)) = f (x, y), M -a.e., for all f ∈ L2 (E, M ). Then for any g ∈ L2 (X, µ), if ( g(x), if x = y, f (x, y) = 0, otherwise, then f ∈ L2 (E, M ) and g(T −1 (x)) = f (T −1 (x), T −1 (x)) = f (x, x) = g(x), µa.e., so T = 1.) Now notice that UTn ×Tn → 1 in U (L2 (E, M )) ⇒ Tn → 1 in N [E]. This is because for any T ∈ N [E], S ∈ [E], if f is the characteristic function of
6. AUTOMORPHISM GROUPS OF EQUIVALENCE RELATIONS
43
graph(S) ⊆ E, then δu (T ST −1 , S) ≤ kUT ×T (f ) − f k22 . On the other hand T 7→ UT ×T is clearly a Borel homomorphism from the Polish group N [E] into the Polish group U (L2 (E, M )), thus it is continuous, so if Tn → 1 in N [E], then UTn ×Tn → 1 in U (L2 (X, M )). So T 7→ UT ×T is a homeomorphism of N [E] with a (necessarily closed) subgroup of U (L2 (X, M )), and we have the following fact. Proposition 6.3. For aperiodic E, the map T 7→ UT ×T from N [E] to U (L2 (E, M )) is a (topological group) isomorphism of N [E] with a closed subgroup of U (L2 (E, M )). Thus we can identify each T ∈ N [E] with the corresponding unitary operator UT ×T on L2 (E, M ). (C) It follows from 4.1 that for every ergodic E, N [E] can be also identified with the automorphism group, Aut([E]), of the (abstract) group [E]. More precisely, every S ∈ N [E] gives rise to the automorphism T 7→ ST S −1 of [E] and every automorphism of the (abstract) group [E] is of this form for a unique S ∈ N [E]. Next notice that [E] is the smallest non-trivial normal subgroup of N [E]. Indeed, if K N [E] is another normal subgroup, then, since [E] is simple (by 4.6), either [E] ≤ K or else [E] ∩ K = {1}. In the latter case K ⊆ C[E] , the centralizer of [E] in Aut(X, µ), which by 4.11 is trivial, so K = {1}. Assume now E, F are ergodic and f : N [E] → N [F ] is an (abstract) group isomorphism. Then f sends [E] onto [F ] so, by 4.1, E ∼ = F and there is ϕ ∈ Aut(X, µ) such that f (T ) = fϕ (T ) = ϕT ϕ−1 , ∀T ∈ [E]. Thus fϕ maps [E] onto [F ] and therefore fϕ maps N [E] onto N [F ]. Consider f −1 ◦fϕ . This is an automorphism of N [E] which is trivial on [E]. Since [E] is simple, nonabelian, a theorem of Burnside (see Thomas [Tho], 1.2.8) asserts that every automorphism of N [E] (which can be identified with the automorphism group of [E]) is inner and thus there is ψ ∈ N [E] such that f −1 ◦ fϕ (T ) = ψT ψ −1 , ∀T ∈ N [E], and as f −1 ◦ fϕ (T ) = T for T ∈ [E], ψT ψ −1 = T, ∀T ∈ [E], i.e., ψ ∈ C[E] , so ψ = 1. Thus f (T ) = fϕ (T ), ∀T ∈ N [E]. So we have the following Reconstruction Theorem for N [E]. Theorem 6.4. If E, F are ergodic equivalence relations, the following are equivalent: (i) E ∼ = F, (ii) N [E], N [F ] are isomorphic as abstract groups. Moreover, for any algebraic isomorphism f : N [E] → N [F ] there is unique ϕ ∈ Aut(X, µ) with f (T ) = ϕT ϕ−1 , ∀T ∈ N [E]. (D) Finally, we note that Danilenko [Da] has shown that (as a topological group) N [E] is contractible, when E is ergodic, hyperfinite. We will see some further properties of such N [E] in the next section. Comments. For the definition of the topology on N [E], see HamachiOsikawa [HO], Gefter [Ge], Danilenko [Da].
44
I. MEASURE PRESERVING AUTOMORPHISMS
7. The outer automorphism group (A) The group [E] is normal in N [E] and the quotient Out(E) = N [E]/[E] is called the outer automorphism group of E. If E is aperiodic and N [E] has the Polish topology τN [E] , then Out(E) with the quotient topology will be a Polish group iff [E] is closed in N [E]. The group [E] may or may not be closed in N [E] (we will say more about this below) but it is always a Polishable subgroup of N [E] with the corresponding Polish topology being equal to the uniform topology on [E]. This is because τN [E] |[E] ⊆ u|[E]. Moreover we have that [E] will be closed in N [E] exactly when τN [E] |[E] = u|[E], i.e., for Tn ∈ [E], Tn STn−1 → S uniformly, ∀S ∈ [E], implies Tn → 1 uniformly. If E = EΓX , this is also equivalent to the assertion that Tn → 1 weakly and ∀γ ∈ Γ(Tn γTn−1 → γ uniformly) implies Tn → 1 uniformly. Proposition 7.1. If E is ergodic and {gn } is dense in ([E], u), then [ [E] = N [E] ∩ {T : δu (T, gn ) < 1}. n
In particular, [E] is open in (N [E]), u) and an Fσ Polishable subgroup of (N [E], τN [E] ). Proof. If T ∈ N [E] and δu (T, gn ) < 1, then T (x) = gn (x) and thus T (x)Ex on a positive measure set. Then, by the ergodicity of E, T (x)Ex, a.e., i.e., T ∈ [E]. Since every open ball in u is Fσ in w and w|N [E] ⊆ τN [E] , it follows that [E] is Fσ in (N [E], τN [E] ). 2 The preceding result should be contrasted with the fact that, for ergodic E, [E] is Π03 -complete in (Aut(X, µ), w) (see the paragraph after 3.1). (B) We next study the complexity of Out(E), when E is ergodic, hyperfinite. Theorem 7.2 (Hamachi-Osikawa [HO]). If E is ergodic, hyperfinite, then [E] is dense in N [E]. Proof. Take E = E0 on 2N . Consider the basic open sets {Ns }s∈2n and let Gn be the permutation group of 2n . We view each g ∈ Gn as a member of [E0 ] (a dyadic S permutation) by letting g(sˆx) = g(s)ˆx. So G1 ⊆ G2 ⊆ . . . and G∞ = n Gn is dense in ([E0 ], u). Finally note that if for s ∈ 2n , gs is the transposition gs = (0n s) that switches 0n , s (with g0n = id), then Gn is generated by {gs }s∈2n (as (a b) = (x a)(x b)(x a)). If T ∈ N [E0 ], n ≥ 1, we will find Tn ∈ [E0 ] such that Tn (Ns ) = T (Ns ), Tn gs Tn−1 = T gs T −1 , for any s ∈ 2n . Then Tn gTn−1 = T gT −1 for any g ∈ Gn , so Tn gTn−1 → T gT −1 uniformly for any g ∈ G∞ , and Tn → T weakly, thus Tn → T in N [E].
7. THE OUTER AUTOMORPHISM GROUP
45
Definition of Tn : Let R ∈ [E0 ] be such that R(N0n ) = T (N0n ). Then define Tn on Ns , s ∈ 2n , as follows: Tn (x) = T gs T −1 Rgs (x).
(*)
As gs , T gs T −1 , R ∈ [E0 ], clearly Tn ∈ [E0 ]. Also Tn |N0n = R|N0n , Tn (Ns ) = T (Ns ). We now check that Tn gs Tn−1 (y) = T gs T −1 (y), ∀s ∈ 2n , ∀y. If y 6∈ T (Ns ) ∪ T (N0n ), T gs T −1 (y) = T T −1 (y) = y and Tn gs Tn−1 (y) = Tn Tn−1 (y) = y. If y ∈ T (Ns ) ∪ T (N0n ) this follows from (∗ ). 2 Since clearly [E0 ] 6= N [E0 ] (e.g., any infinite A ⊆ N induces an element of N [E0 ] \ [E0 ] via x 7→ x + χA (mod 2)), it follows that for any ergodic, hyperfinite E, Out(E) is not Polish. In fact one can say quite a bit more. Call Out(E) turbulent if the action of ([E], u) by (right) translation on N [E] is turbulent. In that case, the equivalence relation induced by the cosets of [E] in N [E] admits no classification by countable structures (i.e., Out(E) cannot be Borel embedded in the isomorphism types of countable structures). Theorem 7.3 (Kechris). If E is not E0 -ergodic, then every element of [E] (closure of [E] in N [E]) is turbulent for the translation action of ([E], u) on N [E]. In particular, if also [E] is not closed in N [E], the translation action of ([E], u) on [E] is turbulent and the associated coset equivalence relation admits no classification by countable structures. Corollary 7.4. If E is ergodic, hyperfinite, then Out(E) is turbulent. Proof. Let G = [E]. Since the periodic g ∈ [E] are dense in ([E], u), fix a countable set {gj } of periodic elements in [E] which is dense in ([E], u). Then a basic nbhd of g ∈ G has the form n \ U= {S ∈ G : δu0 (Sgj S −1 , ggj g −1 ) < }. j=1
Also fix a basic nbhd V of 1 in ([E], u). We will show that O(g, U, V ) is dense in U . Now {gT −1 : T ∈ [E]} is dense in G, so {gT −1 ∈ U : T ∈ [E]} is dense in U . Therefore {ggj−1 ∈ U : j = 0, 1, . . . } is dense in U , thus it is enough, for each ρ > 0, and h ∈ {gj } such that gh−1 ∈ U , to show that O(g, U, V ) ∩ {S : δu0 (S, gh−1 ) < ρ} = 6 ∅. Note that for S ∈ G: gS ∈ U ⇔ ∀j ≤ n[δu0 (gSgj S −1 g −1 , ggj g −1 ) < ] ⇔ ∀j ≤ n[δu0 (Sgj S −1 , gj ) < ]. It is thus enough, for each ρ > 0, to find a continuous path λ 7→ hλ in 0 ([E], u) with h0 = 1 such that δu0 (h−1 λ gj hλ , gj ) < , ∀j ≤ n, and δu (h, h1 ) < ρ. 0 −1 Let α < be such that δu (h gj h, gj ) < α, ∀j ≤ n. Using the notation of 5.2, ¯ g¯j , j ≤ n, which are in [π −1 (EN ) ∩ E] for some large enough we can find h, ¯ < δ0 , δ 0 (gj , g¯j ) < δ0 , ∀j ≤ n, where 0 < δ0 < ρ, α + 6δ0 < . N and δu0 (h, h) u 0 −1 ¯ ¯ g¯j ) < 3δ0 + α + δ0 = α + 4δ0 . Then, as in the proof of Then δu ((h) g¯j h,
46
I. MEASURE PRESERVING AUTOMORPHISMS
¯ 5.2, there is a continuous path λ 7→ hλ in ([E], u) such that h0 = 1, h1 = h −1 0 0 0 ¯ < δ0 < and δu (hλ g¯j hλ , g¯j ) < α + 4δ0 , ∀j ≤ n. Then δu (h, h1 ) = δu (h, h) −1 0 0 ρ, δu (hλ gj hλ , gj ) < 2δu (gj , g¯j ) + α + 4δ0 ≤ α + 6δ0 < . 2 Remark (Tsankov). As we noted earlier, N [E] is closed in the space (Aut(X, µ), u) and so δu is a complete metric on N [E]. Also as we saw in the proof of 7.1 δu (S, T ) = 1 if S ∈ [E], T ∈ N [E] \ [E], so for any two distinct cosets [E]S 6= [E]T of [E] in N [E], δu (S 0 , T 0 ) = 1, ∀S 0 ∈ [E]S, ∀T 0 ∈ [E]T . It follows that (N [E], u) is separable (i.e., Polish) iff Out(E) is countable iff τN [E] = u|N [E]. Remark. Consider the shift action of a countable group Γ on (X Γ , µΓ ), given by γ · p(δ) = p(γ −1 δ), with corresponding equivalence relation E. Any S ∈ Aut(X, µ) induces S ∗ ∈ Aut(X Γ , µΓ ) given by S ∗ (p) = (γ 7→ S(p(γ)). Clearly S ∗ ∈ N [E], in fact S ∗ γ = γS ∗ , ∀γ ∈ Γ. It is easy to check that δu (S ∗ , T ) = 1, for every S 6= 1, T ∈ [E]. It follows that S 7→ S ∗ is a Borel, therefore, continuous embedding of (Aut(X, µ), w) into N [E] with image having trivial intersection with [E]. In fact, it is a homeomorphism of (Aut(X, µ), w) with (a necessarily closed) subgroup of N [E] since, in view of 6.2, Sn∗ → S ∗ in N [E] iff Sn∗ → S ∗ weakly and this is easily equivalent to Sn → S weakly. If π : N [E] → Out(E) is the projection map and G = {S ∗ : S ∈ Aut(X, µ)}, then it is easy to check that π|G is a homeomorphism, so (Aut(X, µ), w) embeds topologically as a subgroup of Out(E). So, if moreover Out(E) is Polish, this copy of (Aut(X, µ), w) in Out(E) is closed, i.e., [E]G is closed in N [E]. Finally note that for any distinct T1 , T2 ∈ Aut(X, µ), δu (T1∗ , T2∗ ) = 1, i.e., G is discrete in u. In particular, taking Γ = Z, we see that if E is ergodic, hyperfinite, then N [E] contains a closed subgroup G isomorphic to (Aut(X, µ), w) with G ∩ [E] = {1} and moreover G is discrete in u. Remark. Assume now that E is ergodic, hyperfinite. Then Connes and Krieger [CK] have shown that Out(E) has the following very strong property: Every two elements of Out(E) which have the same order (finite or infinite) are conjugate. Since, by the preceding remark, Out(E) contains a copy of Aut(X, µ), which is simple and has elements of any order, it follows that Out(E) is simple. Also since in Aut(X, µ) every element is a commutator and a product of three involutions (see Section 2, (D)) the same is true for Out(E). Going up to N [E] we see that [E] is the unique non-trivial normal subgroup of N [E]. Moreover, using again the preceding remark and the result of Connes-Krieger, we see that there is a copy G of Aut(X, µ) in N [E] so that for every T ∈ N [E] some conjugate of T is of the form S1 S2 , where S1 ∈ [E] and S2 ∈ G. It follows (using also 2.10 and the comments in Section 4) that every element of N [E] is the product of 2 commutators and 6 involutions. I do not know if every (algebraic) automorphism of Out(E)
7. THE OUTER AUTOMORPHISM GROUP
47
is inner. Also I do not know whether the following is true: If F is ergodic and Out(F ) is algebraically isomorphic to Out(E), then F is hyperfinite. Bezuglyi and Golodets [BG2] have generalized the Connes-Krieger theorem to show that if π1 , π2 : G → N [E] are two homomorphisms of a countable amenable group into N [E] and π i : G → Out(E) is defined as π i = ϕ◦πi , with ϕ : N [E] → Out(E) the canonical surjection, then π 1 , π 2 are conjugate (i.e., ∃θ ∈ Out(E) with π 1 (g) = θπ 2 (g)θ−1 , ∀g ∈ G) iff ker(π1 ) = ker(π2 ). Remark. Consider now a Borel measure preserving action of a countable group Γ on (X, µ) and the associated equivalence relation EΓX = E. We denote by CΓ the stabilizer of the action, i.e., the set of T ∈ Aut(X, µ) such that (writing γ(x) = γ · x) T γT −1 = γ, ∀γ ∈ Γ. This is equivalent to T (γ · x) = γ · T (x), ∀γ ∈ Γ, i.e., T ∈ CΓ iff T is an isomorphism of the action. In particular, CΓ ≤ N [E]. By Proposition 6.2, assuming E is aperiodic, τN [E] |CΓ = w|CΓ . Thus, since CΓ is obviously a closed subgroup of (Aut(X, µ), w), and therefore Polish, CΓ is a closed subgroup of (N [E], τN [E] ). Now CΓ ≤ N [E] acts by conjugation on [E]. Since (CΓ , w) = (CΓ , τN [E] ) is a closed subgroup of Iso([E], δu ), when we identify T ∈ CΓ with S ∈ [E] 7→ T ST −1 ∈ [E], and the evaluation action of Iso([E], δu ) (with the pointwise convergence topology) on ([E], u) is continuous, it follows that the conjugation action of (CΓ , w) on ([E], u) is continuous. Consider then the semidirect product CΓ n [E] (for the conjugation action), which we take here to be the space CΓ × [E] with the product topology and multiplication defined by (T1 , S1 )(T2 , S2 ) = (T1 T2 , (T2−1 S1 T2 )S2 ) (see Appendix I, (B)). This is again a Polish group. Let ϕ : CΓ n [E] → N [E] be defined by ϕ(T, S) = T S. This is a continuous homomorphism of the Polish group CΓ n E into the group (N [E], τN [E] ) whose range is the group CΓ [E] generated by CΓ and [E]. In particular, CΓ [E] is Polishable. In the case CΓ ∩ [E] = {1} (and this is quite common, see 14.6 below), then ϕ is a continuous injective homomorphism of CΓ n [E] onto CΓ [E]. Thus if additionally CΓ [E] is closed, ϕ is a topological group isomorphism of CΓ n [E] with (CΓ [E], τN [E] ). Since ϕ({1} × [E]) = [E], we conclude that [E] is closed in N [E]. Thus (CΓ ∩ [E] = {1} and CΓ [E] is closed in N [E]) ⇒ [E] is closed in N [E]. Comments. Very interesting explicit calculations of the outer automorphism groups of equivalence relations induced by certain actions have been
48
I. MEASURE PRESERVING AUTOMORPHISMS
obtained by Gefter [Ge], Gefter-Golodets [GG], Furman [Fu1], Popa [Po4], Ioana-Peterson-Popa [JPP] and Popa-Vaes [PV]. Among them are examples where Out(E) is trivial, i.e., N [E] = [E]. 8. Costs and the outer automorphism group (A) We establish here a connection between the cost of an equivalence relation and the structure of its outer automorphism group. Concerning the theory of costs, see Gaboriau [Ga]. We follow in this section the terminology and notation of Kechris-Miller [KM], Ch. 3. In particular, the cost of an equivalence relation E on (X, µ) is denoted by Cµ (E). Theorem 8.1 (Kechris). If the outer automorphism group of an ergodic equivalence relation E is not Polish (i.e., if [E] is not closed in N [E]), then Cµ (E) = 1. Proof. To say that [E] is not closed in N [E] means that the identity map from ([E], u) into N [E] is not a homeomorphism, i.e., the identity map from ([E], τN [E] ) into ([E], u) is not continuous, or, equivalently, that there is a sequence {Tn } ⊆ [E] such that δu (Tn T, T Tn ) → 0, ∀T ∈ [E], but δu (Tn , 1) 6→ 0. Call a sequence {Tn } ⊆ [E] good if ∃ > 0∀n(δu (Tn , 1) ≥ ) & ∀T ∈ [E](δu (Tn T, T Tn ) → 0). Thus [E] is not closed in N [E] iff there is a good sequence. Note also that a subsequence of a good sequence is good. Lemma 8.2. If {Tn } is good, then EhTn in (= the equivalence relation induced by {Tn }) is aperiodic. We will assume this and complete the proof. Lemma 8.3. If {Tn } is good, S ∈ [E] is aperiodic, and > 0, then there is a sequence i0 < i1 < i2 < . . . such that Cµ (EhS,Ti i` ) < 1 + for any k` subsequence ik0 < ik1 < ik2 < . . . . Proof. We first show that for any δ > 0 there is i(δ) such that for any i ≥ i(δ) there is Ai with µ(Ai ) < δ and {S, Ti |Ai } is an L-graphing of EhS,Ti i . Since S is aperiodic, we can find a complete Borel section A of ES with µ(A) < δ/3. Then we can find large N so that µ(B) < δ/3, where B = {x ∈ X : ∀|n| ≤ N (S n (x) 6∈ A)}. Next fix i(δ) such that for all i ≥ i(δ) and |n| ≤ N , δu (Ti S n , S n Ti ) = δu (S −n Ti S n , Ti ) <
δ . 3(2N + 1)
For such i, if Ci = {x : ∃|n| ≤ N (S −n Ti S n (x) 6= Ti (x))}, then µ(Ci ) < δ/3.
8. COSTS AND THE OUTER AUTOMORPHISM GROUP
49
If x 6∈ Ci , x 6∈ B, there is |n| ≤ N with S n (x) ∈ A and Ti (x) = n i S (x). Thus {S, Ti |(A ∪ B ∪ Ci )} forms an L-graphing of EhS,Ti i and if Ai = A ∪ B ∪ Ci , then µ(Ai ) < δ. It follows that we can find i0 < i1 < · · · < ik < ik+1 < . . . such that i ≥ ik ⇒ ∃Bi (µ(Bi ) < k+1 and {S, Ti |Bi } is an L-graphing of EhS,Ti i ). 2 Let ik0 < ik1 < . . . be a subsequence of {ik }. Then ik` ≥ i` , so we can find D` with µ(D` ) < 2`+1 and {S, Tik` |D` } is an L-graphing of EhS,Ti i , so that S −n T
k`
{S, Tik` |D` }` is an L-graphing of EhS,Ti i` and therefore Cµ (EhS,Ti i` ) < k` k` P 1+ ∞ = 1 + . 2 `+1 `=0 2 Lemma 8.4. If [E] is not closed in N [E] and there are aperiodic transformations S1 , . . . , Sk ∈ [E] with E = EhS1 ,...,Sk i , then Cµ (E) = 1. Proof. Fix > 0. Repeatedly applying the preceding lemma, we can find a good sequence {Tn } such that Cµ (EhSi ,Tn in ) ≤ 1 + , ∀i ≤ k. k Since, by Lemma 8.2, EhTn in is aperiodic, it follows (see Kechris-Miller [KM], 23.5) that Cµ (E) − 1 = Cµ (
k _
EhSi ,Tn in ) − 1
i=1
≤
k X
[Cµ (EhSi ,Tn in ) − 1] ≤ .
i=1
Since is arbitrary, Cµ (E) = 1.
2
The next fact is true in the pure Borel category (where we interpret [E] as consisting of all Borel automorphisms T with T (x)Ex, for all x). Lemma 8.5. Let E be a countable aperiodic Borel equivalence relation and S ∈ [E]. Then there is an aperiodic T ∈ [E] with ES ⊆ ET . Proof. We can assume that every orbit of S is finite. Let then A be a Borel transversal for ES . Then E|A is aperiodic, so (by the proof of 3.5, (i)) let F ⊆ E|A be aperiodic hyperfinite. Then F ∨ ES is aperiodic, hyperfinite and so of the form ET for some aperiodic T ∈ [E]. 2 Now, to complete the proof, assume E is as in the statement of the theorem. By 3.5 there is U0 ∈ [E] which is ergodic (and thus aperiodic). By 8.3, there is a good sequence {Tn } such that Cµ (EhU0 ,Tn in ) < 2. Let E = EhU1 ,U2 ,... i , Ui ∈ [E]. Then if Em = EhU0 ,...,Um ,Tn in , E1 ⊆ E2 ⊆ . . . and each Em is ergodic. Also Cµ (Em ) ≤ m + Cµ (EhU0 ,Tn in ) < m + 2 < ∞,
50
I. MEASURE PRESERVING AUTOMORPHISMS
so (see Kechris-Miller [KM], 27.7) Em = EhS1 ,...,Sk i for some S1 , . . . , Sk ∈ [Em ], which by 8.5 we can assume are aperiodic. But [Em ] is not closed in S N [Em ] since {Tn } is good for Em as well. So, by 8.4, Cµ (Em ) = 1. As m Em = E, it follows (see Kechris-Miller [KM], 23.5) that Cµ (E) = 1. So it only remains to give the proof of 8.2. Proof of Lemma 8.2. I will give an argument due to Ben Miller which is simpler than my original one. Assume, towards a contradiction, that there is a set of positive measure on which F = EhTn in is periodic. It follows that there is a partition X = A ∪ B into two F -invariant Borel sets of positive measure with F |A periodic. By ergodicity of E, we can assume that A, B are complete sections for E. S Let {gn } be a sequence of Borel involutions such that E = n graph(gn ). Also let {An } be a sequence of partial Borel F -transversals (i.e., each An S meets every F -class in at most one point) such that n An = A. Let finally Gn , Hn be pairwise disjoint such that Gn ∪Hn = supp(gn ) and gn (Gn ) = Hn . Put 0 G0n,m = {x ∈ Gn ∩ Am : gn (x) ∈ B}, Hn,m = G0n,m ∪ gn (G0n,m )
and let 0 0 hn,m = gn |Hn,m ∪ id|(X \ Hn,m ).
Then {hn,m } is a family of involutions and xEy & x ∈ A & y ∈ B ⇔ ∃n, m(hn,m (x) = y & x 6= y). S Also hn,m ∈ [E] and n,m supp(hn,m ) = X. Claim. If z ∈ supp(hn,m ) and Tk hn,m Tk−1 (z) = hn,m (z), then Tk (z) = z. Proof of the claim. Assume z ∈ supp(hn,m ) and Tk hn,m Tk−1 (z) = hn,m (z) but Tk (z) 6= z (and thus Tk−1 (z) 6= z as well). Case (1). z ∈ A. Either Tk−1 (z) 6∈ supp(hn,m ), so Tk hn,m Tk−1 (z) = z = hn,m (z), which is impossible or Tk−1 (z) ∈ supp(hn,m ), so z 6= Tk−1 (z) are F -equivalent and in supp(hn,m ) ∩ A ⊆ Am , which is an F -transversal, a contradiction. Case (2). z ∈ B. Let hn,m (z) = y, so y ∈ A and hn,m (y) = z. As in Case (1), Tk−1 (z) ∈ supp(hn,m ). Put hn,m (Tk−1 (z)) = w. Then wF y and w 6= y. But also w, y ∈ supp(hn,m ) ∩ A ⊆ Am , a contradiction. So we see that supp(hn,m ) ∩ {z : Tk (z) 6= z} ⊆ {z : Tk hn,m Tk−1 (z) 6= hn,m (z)}. Let An,m ⊆ supp(hn,m ) be pairwise disjoint with [ [ An,m = supp(hn,m ) = X. n,m
n,m
8. COSTS AND THE OUTER AUTOMORPHISM GROUP
51
Then {z ∈ An,m : Tk (z) 6= z} ⊆ {z ∈ An,m : Tk hn,m Tk−1 (z) 6= hn,m (z)}, thus δu (Tk , 1) ≤
X
µ(An,m ∩ {z : Tk hn,m Tk−1 (z) 6= hn,m (z)}).
n,m
Fix > 0 and find a finite set F ⊆ N2 such that X µ(An,m ) < (n,m)6∈F
and then K0 large enough so that for all (n, m) ∈ F , k ≥ K0 ⇒ µ({z : Tk hn,m Tk−1 (z) 6= hn,m (z)}) <
. card(F )
Then for such k, δu (Tk , 1) ≤ + card(F ) ·
= 2. card(F )
Thus δu (Tk , 1) → 0, a contradiction.
2
Remark. Note that if δu (Tn , 1) → 0, then there is a subsequence {Tni } such that EhTni ii is not aperiodic (in fact it is equality on a set of posiP tive measure). To Psee this, find {ni } so that i δu (Tni , 1) < ∞. Then if Bni = supp(Tni ), i µ(Bni ) < ∞, so, by Borel-Cantelli, µ({x : ∀i∃j ≥ i(x ∈ Bnj )}) = 0, or (neglecting as usual null sets) ∀x∃i∀j ≥ i(Tnj (x) = x). Then for some set B of positive measure and some i, x ∈ B ⇒ ∀j ≥ i(Tnj (x) = x), i.e., EhTnj ij≥i |B = equality|B. Thus we see that, given any sequence {Tn } ⊆ E with Tn → 1 in N [E], the following are equivalent: (i) ∃ > 0∀n(δu (Tn , 1) ≥ ) (i.e., {Tn } is good), (ii) For every subsequence {ni }, EhTni ii is aperiodic. Remark. By definition, [E] not closed in N [E] means that there is a sequence {Tn } ⊆ [E] with δu (Tn , 1) 6→ 0 but Tn → 1 in N [E]. We note here that if E is actually E0 -ergodic, then, if [E] is not closed in N [E], there is such {Tn } with δu (Tn , 1) = 1, i.e., supp(Tn ) = X. Indeed we have some {Sn } ⊆ [E] and > 0 with δu (Sn , 1) ≥ and Sn → 1 in N [E]. Let An = {x : Sn (x) = x}, so that µ(X \ An ) ≥ . If δu (Sn , 1) 6→ 1, then we can assume that for some δ > 0, δu (Sn , 1) ≤ 1 − δ, so µ(An ) ≥ δ. Thus µ(An )(1 − µ(An )) 6→ 0. Fix T ∈ [E]. If Bn = {x : T −1 Sn (x) 6= Sn T −1 (x)},
52
I. MEASURE PRESERVING AUTOMORPHISMS
then µ(Bn ) → 0 (as Sn → 1 in N [E]). If x 6∈ Bn , then x ∈ An ⇒ T −1 Sn (x) = T −1 (x) = Sn (T −1 (x)) ⇒ T −1 (x) ∈ An ⇒ x ∈ T (An ), so An \T (An ) ⊆ Bn and therefore µ(An \T (An )) → 0. Thus µ(An ∆T (An )) = 2µ(An \ T (An )) → 0. It follows that E is not E0 -ergodic, a contradiction. Thus δu (Sn , 1) → 1. Let Un ∈ [E] be an involution with support An and put ( Sn (x), if x 6∈ An , Tn (x) = Un (x), if x ∈ An . Then Tn ∈ [E], δu (Sn , Tn ) ≤ µ(An ) → 0 and δu (Tn , 1) = 1. Also clearly Tn → 1 in N [E]. Remark. The converse of 8.1 is not true. By 9, (A) and 9.1 below there are groups Γ such that for any measure preserving, free, ergodic action of Γ, we have [EΓX ] closed in N [EΓX ] but Cµ (EΓX ) = 1. For example, one can take any ICC (infinite conjugacy classes) group with property (T), fixed price and cost equal to 1, for instance SL3 (Z). (B) The preceding result 8.1 can be used to give a partial answer to a question of Jones-Schmidt [JS], p. 113 and Schmidt [Sc2], 4.6. They ask for a characterization of when [E] is closed in N [E]. The following corollary resolves this when E is treeable, i.e., there is a Borel acyclic graph whose connected components are the E-classes. Corollary 8.6. Let E be an ergodic equivalence relation. Then if E is treeable, the following are equivalent: (i) [E] is closed in N [E]. (ii) E is not hyperfinite. Proof. We have already seen that if E is hyperfinite, [E] is dense in (and not equal to) N [E]. Conversely, if [E] is not closed, then by 8.1, Cµ (E) = 1, so, as E is treeable, E is hyperfinite (see Kechris-Miller [KM], 22.2 and 27.10). 2 It is interesting to note here that although the statement of this result does not involve costs, the proof given here makes use of this concept. For further results concerning the question of when [E] is closed in N [E], see the next section and Section 29, (C). 9. Inner amenability (A) Recall that a countable group Γ is inner amenable if there is a mean on Γ \ {1} invariant under the conjugacy action of Γ or equivalently there is a finitely additive probability measure (f.a.m.) on Γ\{1} invariant under the conjugacy action of Γ. See Effros [Ef1], where this notion was introduced, and the survey Bedos-de la Harpe [BdlH]. Examples of such groups are: (i) amenable groups,
9. INNER AMENABILITY
53
(ii) groups that have a finite conjugacy class 6= {1}, (iii) direct products Γ × ∆, where Γ 6= {1} is amenable, (iv) weakly commutative groups (i.e., groups Γ such that for every finite F ⊆ Γ there is a γ ∈ Γ \ {1} commuting with each element of F ). For example, direct sums Γ1 ⊕ Γ2 ⊕ . . . of countable non-trivial groups, (v) the Thompson group F = hx0 , x1 , x2 , . . . |x−1 i xn xi = xn+1 , i < ni. On the other hand the free groups Fn , n ≥ 2, are not inner amenable. Also ICC groups with property (T) are not inner amenable. (B) The following result was proved in Jones-Schmidt [JS]. Theorem 9.1 (Jones-Schmidt). Given a free, measure preserving action of a countable group Γ on (X, µ), if [EΓX ] is not closed in N [EΓX ], then Γ is inner amenable. Proof. Since, letting E = EΓX , [E] is not closed in N [E], there is a sequence Tn ∈ [E] with δu (Tn , 1) ≥ , some > 0, and Tn γTn−1 → γ uniformly, ∀γ ∈ Γ. Put Tn (x) = α(n, x) · x, where α(n, x) ∈ Γ. Let U be a non-principal ultrafilter on N and define a positive linear functional on `∞ (Γ) by Z ϕ(f ) = lim f (α(n, x))dµ(x). n→U
Note that if χ1 is the characteristic function of {1}, then ϕ(χ1 ) 6= 1 as R ϕ(χ1 ) = limn→U R χ1 (α(n, x))dµ(x) = limn→U {x:Tn (x)=x} dµ(x) = limn→U µ({x : Tn (x) = x}) = limn→U (1 − δu (Tn , 1)) ≤ 1 − < 1. We will next see that ϕ is conjugacy invariant. Then clearly ψ(f ) =
ϕ(f ∗ ) , 1 − ϕ(χ1 )
where f ∗ (γ) = f (γ) if γ 6= 1, f ∗ (1) = 0, is a conjugacy invariant mean on Γ \ {1}, so we are done. For γ ∈ Γ, f ∈ `∞ (Γ), put (γ · f )(δ) = f (γ −1 δγ). We need to verify that ϕ(γ · f ) = ϕ(f ). We have Z ϕ(γ · f ) = lim (γ · f )(α(n, x))dµ(x) n→U Z = lim f (γ −1 α(n, x)γ)dµ(x). n→U
Let An = {x : Tn then
(γ −1
· x)) 6= γ −1 · Tn (x)}, so that µ(An ) → 0. If x 6∈ An , Tn (γ −1 · x) = γ −1 · Tn (x),
54
I. MEASURE PRESERVING AUTOMORPHISMS
so α(n, γ −1 · x)γ −1 · x = γ −1 α(n, x) · x and, as the action is free, α(n, γ −1 · x)γ −1 = γ −1 α(n, x), so γ −1 α(n, x)γ = α(n, γ −1 · x). Thus Z ϕ(γ · f ) = lim
n→U
f (γ
−1
Z
f (α(n, γ −1 · x))dµ(x).
α(n, x)γ)dµ(x) + lim
n→U
An
X\An
The second summand is equal to Z Z lim f (α(n, x))dµ(x) = ϕ(f ) − lim n→U
n→U
γ −1 ·(X\An )
f (α(n, x))dµ(x).
γ −1 ·An
Thus Z ϕ(γ · f ) − ϕ(f ) = lim
n→U
f (γ −1 α(n, x)γ)dµ(x)−
ZAn
lim
n→U
f (α(n, x))dµ(x), γ −1 ·An
which equals 0, since µ(An ) → 0.
2
It follows from 9.1 that for every free, measure preserving action of a countable group Γ, which is ICC and has property (T), [EΓX ] is closed in N [EΓX ] (see Gefter-Golodets [GG]). (C) The converse of 9.1 is literally false as stated, in view of the following simple fact. Proposition 9.2. If Γ = ∆ × F2 , where ∆ is non-trivial finite, then for every free, measure preserving, ergodic action on (X, µ), [EΓX ] is closed in N [EΓX ]. Proof. This follows from 8.1, as the cost C(Γ) is greater than 1 and Γ has fixed price but one can also give a direct proof. Let E = EΓX . Let Y = X/∆ (we view here ∆ as a subgroup of Γ and similarly for F2 ). Let π : X → Y be the canonical projection and put ν = π∗ µ. Then F2 acts on Y by γ ·(∆·x) = ∆·(γ ·x). This is a free, measure preserving, ergodic action of F2 on Y . If [EΓX ] is not closed in N [EΓX ], we will show that [EFY2 ] is not closed in N [EFY2 ], contradicting 9.1. Let Tn ∈ [EΓX ] be such that δu (Tn , 1) 6→ 0 but Tn → 1 in N [EΓX ]. Thus if An = {x : ∀δ ∈ ∆(Tn (δ · x) = δ · Tn (x))}, then µ(An ) → 1. Then for every ∆ · x ∈ Y such that ∆ · x ∩ An 6= ∅ define T˜n (∆ · x) = ∆ · Tn (y), for y ∈ ∆ · x ∩ An .
9. INNER AMENABILITY
55
This is well-defined. Clearly T˜n (∆ · x)EFY2 ∆ · x and T˜n is 1-1 and measure preserving on its domain, which has ν-measure equal to µ(∆ · An ) → 1. So we can extend T˜n to Sn ∈ [EFY2 ]. We will check that {Sn } witnesses that [EFY2 ] is not closed in N [EFY2 ]. For this it is enough to verify: (i) Sn → 1 weakly, (ii) Sn γSn−1 → γ uniformly, ∀γ ∈ F2 , (iii) δu (Sn , 1) 6→ 0. Now (i), (ii) easily follow from the fact that {Tn } has the same properties. To prove (iii) we will use the following simple fact: If E is a smooth equivalence relation and Un ∈ [E], Un → 1 weakly, then Un → 1 uniformly. This S is because w and u agree on [E], as [E] is weakly closed (or write X = i Ai , where {Ai } is a pairwise disjoint family of partial Borel transversals and note that µ({x ∈ Ai : Un (x) 6= x}) = 12 µ(Un (Ai )∆Ai )). Assume then that (iii) fails, towards a contradiction. From the definition it follows then that X ] such that δ (U , T ) → 0 and so U → 1 weakly (since there is Un ∈ [E∆ u n n n Tn → 1 weakly) and Un 6→ 1 uniformly, a contradiction. 2 Schmidt [Sc2] raised the question of whether there is a converse of 9.1 for ICC groups. Problem 9.3. Assume that Γ is countable, ICC and inner amenable. Is there a free, measure preserving, ergodic action of Γ for which [EΓX ] is not closed in N [EΓX ]? In connection with this question we note the following corollary of 8.1, 8.6. Theorem 9.4 (Kechris). Suppose Γ is a countable group that admits a free, measure preserving, ergodic action on (X, µ) with [EΓX ] not closed in N [EΓX ]. Then C(Γ) = 1. If moreover, Γ is strongly treeable (i.e., every equivalence relation induced by a free, measure preserving action of Γ is treeable), then Γ is amenable. Thus for strongly treeable groups Γ, Γ is amenable iff there is a free, measure preserving, ergodic action of Γ on (X, µ), with [EΓX ] not closed in N [EΓX ]. Note that, in view of 9.4, a positive answer to 9.3 implies that C(Γ) = 1 for any ICC inner amenable group, which is also an open problem. Concerning 9.3, we also have the following partial answer. Proposition 9.5. Let Γ be a countable ICC group which is weakly commutative. Then there is a free, measure preserving, ergodic action of Γ with [EΓX ] not closed in N [EΓX ]. Before we start the proof, let us notice certain general facts about conjugacy shift actions. Let Γ be a countable group, put Γ∗ = Γ \ {1}, ∗
and consider its conjugacy shift action on 2Γ : (γ · p)δ = p(γ −1 δγ).
56
I. MEASURE PRESERVING AUTOMORPHISMS
Proposition 9.6. Let Γ be an infinite countable group. Then Γ is ICC iff ∗ its conjugacy shift action on 2Γ is ergodic iff its conjugacy shift action on ∗ Γ 2 is weak mixing. ∗
∗
Proof. View 2Γ as ZΓ2 = the Cantor group. The conjugacy shift action ∗ ∗ of Γ on ZΓ2 is an action by automorphisms of ZΓ2 and the corresponding d Γ∗ = Z<Γ∗ = P (Γ∗ ) (= the set of finite subsets of Γ∗ ) is the action on Z 2
2
fin
∗ (Γ) = P (Γ∗ )\{∅} and denote by conjugacy action of Γ on Pfin (Γ∗ ). Let Pfin fin ∗ ΓF the stabilizer of F ∈ Pfin (Γ) under this action. Since the trivial character corresponds to ∅, we have that the conjugacy shift action is ergodic iff it is ∗ (Γ)([Γ : Γ ] = ∞) (see, e.g., Kechris [Kec4], 4.3). weak mixing iff ∀F ∈ Pfin F If γ ∈ Γ∗ , then Γ{γ} has infinite index iff the conjugacy class of γ is ∗ (Γ)([Γ : infinite. Also [ΓF : ΓF ∩ Γ{γ} ] < ∞ for any γ ∈ F , thus ∀F ∈ Pfin ∗ ΓF ] = ∞) iff ∀γ ∈ Γ ([Γ : Γ{γ} ] = ∞) iff Γ is ICC. 2
Also note the following fact. Proposition 9.7. If Γ is countable ICC, then the conjugacy shift action of ∗ Γ on 2Γ is free (a.e.). ∗
Proof. Fix γ ∈ Γ∗ and let p ∈ 2Γ be such that γ · p = p, i.e., ∀δ(p(δ) = Thus if hγi acts by conjugation on Γ∗ , p is constant on each orbit. Thus {p : γ · p = p} is null unless all orbits of this action are finite and all but finitely many are trivial, i.e., γ commutes with all but finitely many elements of Γ∗ in which case its conjugacy class is finite, a contradiction. 2 p(γ −1 δγ)).
∗
Proof of 9.5. Let X = 2Γ , µ = the usual product measure and let Γ ∗ act on 2Γ by conjugacy shift. This action is free, measure preserving and ergodic. Let {γn } ⊆ Γ∗ witness the weak commutativity of Γ, i.e., for each δ ∈ Γ, δγn = γn δ for all large enough n. View each γ ∈ Γ as an element of [EΓX ]. (i) γn → 1 weakly: It is enough to show for each set A of the form ∗ A = {p ∈ 2Γ : p(δi ) = ai , i = 1, . . . , k}, where ai ∈ {0, 1}, δ1 , . . . , δk ∈ Γ∗ , that µ(γn (A)∆A) → 0. Now for all large enough n: p ∈ γn (A) ⇔ γn−1 · p ∈ A ⇔ ∀i ≤ k((γn−1 · p)(δi ) = ai ) ⇔ ∀i ≤ k(p(γn δi γn−1 ) = ai ) ⇔ ∀i ≤ k(p(δi ) = ai ) ⇔ p ∈ A. (ii) γn δγn−1 → δ uniformly, ∀δ ∈ Γ: This is clear as γn δγn−1 = δ, if n is large enough. (iii) δu (γn , 1) = 1, as Γ acts freely. 2 (D) Recall from Jones-Schmidt [JS] that a measure preserving action of Γ on (X, µ) is called stable if EΓX (on (X, µ)) is isomorphic to EΓX × E0 (on (X, µ) × (2N , ν), where ν is the usual product measure on 2N ). In
9. INNER AMENABILITY
57
Theorem 3.4 of Jones-Schmidt [JS] this is characterized by the existence of a sequence {Tn } ⊆ [EΓX ] for which we have Tn → 1 in N [EΓX ] but for some non-trivial asymptotically invariant sequence {An } ⊆ MALGµ we have µ(Tn (An )∆An ) 6→ 0. (Here {An } is a non-trivial asymptotically invariant sequence if µ(γ · An ∆An ) → 0, ∀γ ∈ Γ, but µ(An )(1 − µ(An )) 6→ 0.) Note that for such {Tn }, Tn 6→ 1 uniformly, so {Tn } is a witness to the fact that [EΓX ] is not closed in N [EΓX ]. Proposition 9.8. Suppose there exist {γn }, {δn } ⊆ Γ that witness L weak commutativity of Γ but γn δn = 6 δn γn , ∀n. (For example, let Γ = n Γn , Γn ∗ non-abelian.) Then the conjugacy shift action of Γ on 2Γ is stable. ∗
Proof. As in the proof of 9.5, γn → 1 in N [EΓ ]. Put An = {p ∈ 2Γ : ∗ p(δn ) = 1}. Then γn · An = {p ∈ 2Γ : p(γn δn γn−1 ) = 1}, and γn δn γn−1 6= δn , so 1 µ(γn · An ∩ An ) = µ(γn · An )µ(An ) = . 4 Thus µ(γn · An ∆An ) = 21 . Finally notice that µ(γ · An ∆An ) → 0, ∀γ ∈ Γ, as γ · An = An , if n is large enough. 2 We now have the following application, where two actions are orbit equivalent if the induced equivalence relations are isomorphic (see Section 10, (A)). Proposition 9.9. Let {Γn } be a sequence of countable non-trivialL groups at least one of which is not amenable. Then the shift action of Γ = n Γn on ∗ 2Γ is not orbit equivalent to the conjugacy shift action on 2Γ . Proof. Since the shift action of Γ on 2Γ is E0 -ergodic, it is enough to show that the conjugacy shift action is not. Case 1. Eventually the Γn are abelian. Then Γ = ∆ × Z, Z infinite abelian. Consider ∗ 2Γ 3 p 7→ f (p) ∈ 2Z , where f (p)(z) = p(1, z). Then if p, p0 are shift conjugate, clearly f (p) = ∗ f (p0 ). Thus the conjugacy shift action of Γ on 2Γ is not ergodic, as f cannot be constant on a conull set. Case 2. Infinitely many Γn are not abelian, say for n = n1 < n2 < . . . . Choose ani , bni ∈ Γni that do not commute. Let γi = (1, . . . , ani , 1, . . . ), δi = (1, . . . , bni , 1, . . . ). These satisfy the hypothesis of 9.8, so the conjugacy shift ∗ action of Γ on 2Γ is stable, thus not E0 -ergodic. 2 Addendum. Recently Tsankov showed that for any inner amenable ∗ group Γ the conjugacy shift action of Γ on 2Γ is not E0 -ergodic and thus 9.9 is true for any non-amenable, inner amenable group. Kechris and Tsankov [KT] then studied more generally the action of a countable group Γ on a countable set I and its relationship to the corresponding shift action of Γ on 2I . They proved that the following are equivalent: i) The action of Γ on I is amenable (i.e., admits a finitely additive probability
58
I. MEASURE PRESERVING AUTOMORPHISMS
measure), ii) the shift action of Γ on 2I admits non-trivial almost invariant sets (equivalently: admits a non-trivial asymptotically invariant sequence), iii) the Koopman representation associated to the shift action, restricted to the orthogonal of the constant functions, admits non-0 almost invariant vectors (see Section 10, (C) for an explanation of these notions). In particular, if these conditions hold, the shift action of Γ on 2I is not orbit equivalent to the shift action of Γ on 2Γ , provided Γ is not amenable. We finally note the following (almost) characterization of weakly commutative groups. Proposition 9.10. Let Γ be a countable group. i) If Γ is a non-discrete (i.e., not closed) normal subgroup of a Polish group, then Γ is weakly commutative. ii) If Γ is weakly commutative with trivial center, then Γ is a non-discrete normal subgroup of a Polish group. Proof. i) Assume that Γ is a non-discrete normal subgroup of the Polish group G. Consider the map π : G → Aut(Γ), π(g) = (γ 7→ gγg −1 ). This is a Borel homomorphism, so it is continuous, where Aut(Γ) has the pointwise convergence topology. Since Γ is not discrete in G, there is a sequence {γn } ⊆ Γ \ {1} with γn → 1 (in G). So ∀δ∀∞ n(γn δγn−1 = δ), i.e., ∀δ∀∞ n(γn δ = δγn ), where ∀∞ means “for all but finitely many”. Thus Γ is weakly commutative. ii) Consider the map ρ : Γ → Aut(Γ) = G, ρ(γ) = (δ 7→ γδγ −1 ). Since Γ has trivial center, ρ is an isomorphism, so we can identify Γ with ρ(Γ), which is a normal subgroup of G. Let {γn } ⊆ Γ \ {1} witness the weak commutativity of Γ. Then γn 6= 1, γn → 1 in G, so Γ is not discrete in G. 2 Thus a centerless group is weakly commutative iff it is (up to algebraic isomorphism) a non-discrete normal subgroup of a Polish group (or equivalently a dense normal subgroup of a non-discrete Polish group). In the case of groups with non-trivial center (which are automatically weakly commutative) it is not clear if there is a simple characterization of when they are non-discrete normal subgroups of Polish groups. Tsankov pointed out that if Γ has non-trivial finite center and countable automorphism group Aut(Γ), then Γ cannot be a non-discrete normal subgroup of a Polish group G. Otherwise if π : G → Aut(Γ) is defined as in the proof of 9.10, i) and γn are distinct in Γ with γn → 1, then π(γn ) → 1 in Aut(Γ), which is discrete, so π(γn ) = 1 for all large enough n and thus the center of Γ is infinite. For example, one can take Γ = ∆ × Z2 , where ∆ is a simple
9. INNER AMENABILITY
59
countable group which is perfect, i.e., every automorphism is inner. On the other hand, it is well-known that if Γ is infinite abelian, then Γ is a subgroup of a compact abelian Polish group and thus a non-discrete normal ˆ subgroup of a Polish group. Indeed, if Γ is an infinite abelian group and Γ ˆ which separates points (i.e., is its dual, there is a countable subset I ⊆ Γ ∀γ ∈ Γ \ {1}∃ˆ γ ∈ I(ˆ γ (γ) 6= 1)). Then consider the (compact abelian Polish) product group G = TI and embed (algebraically) Γ into G via π(γ) = γ|I ˆ (where we view here γ as a character of Γ). In the reverse direction, it is also not clear what non-discrete Polish groups admit a dense normal subgroup. Clearly every non-discrete abelian Polish group and the infinite symmetric group have this property.
CHAPTER II
The space of actions 10. Basic properties Convention. In the sequel, we consider countable (discrete) groups Γ, unless otherwise explicitly stated. (A) Let Γ be a countable group and (X, µ) a standard probability space. Every measure preserving action a of Γ on (X, µ) gives rise to a homomorphism π : Γ → Aut(X, µ), where π(γ) = γ a with γ a (x) = a(γ, x). Conversely every homomorphism π : Γ → Aut(X, µ) gives rise to a Borel action a of Γ on X such that for each γ ∈ Γ, a(γ, x) = π(γ)(x), a.e. Then identifying, as usual, two actions a, b if they agree almost everywhere (i.e., a(γ, x) = b(γ, x), a.e., ∀γ), we can view the space of measure preserving actions of Γ in (X, µ) as the space of all homomorphisms of Γ into Aut(X, µ), denoted by A(Γ, X, µ) or just A(Γ) if (X, µ) is understood. If Aut(X, µ) is equipped with either the weak or the uniform topology and Aut(X, µ)Γ with the product topology (which we also denote by w or u), then A(Γ, X, µ) is a closed subspace of Aut(X, µ)Γ , so (A(Γ, X, µ), w) is Polish and (A(Γ, X, µ), u) is completely metrizable. Thus an → a weakly (uniformly) iff γ an → γ a weakly (uniformly), ∀γ ∈ Γ. Moreover, if Γ = {γn }, then X δΓ,w (a, b) = 2−n δ¯w (γna , γnb ) n
is a complete compatible metric for w and X δΓ,u (a, b) = 2−n δu (γna , γnb ) n 0 is a compatible complete metric for u (and similarly for δΓ,u defined by the 0 same formula using δu instead of δu ) The group Aut(X, µ) acts on A(Γ, X, µ) by conjugation: (T, a) 7→ T aT −1 , where −1 γ T aT = T γ a T −1 , ∀γ ∈ Γ. 0 This is a continuous action in w or u. Note that δΓ,u , δΓ,u are invariant under conjugation. If a, b are measure preserving actions of Γ on (X, µ), (Y, ν), resp., we say that they are isomorphic, a∼ = b, 61
62
II. THE SPACE OF ACTIONS
if there is a measure preserving bijection ϕ : X → Y such that ϕ ◦ γ a = γ b ◦ ϕ, ∀γ ∈ Γ. Thus two actions a, b ∈ A(Γ, X, µ) are isomorphic iff they are conjugate. We call a, b orbit equivalent, aOEb, if the equivalence relations induced by a, b are isomorphic, i.e., there is a measure preserving bijection ϕ : X → Y such that if Ea = {(x, y) : ∃γ(a(γ, x) = y)}, then xEa y ⇔ ϕ(x)Eb ϕ(y) modulo null sets. We also write this as Ea ∼ = Eb . ∼ Clearly a = b ⇒ aOEb. If a, b ∈ A(Γ, X, µ), let also a ≡ b ⇔ Ea = Eb (modulo null sets). Then clearly OE = (≡)∨(∼ =) and ≡, ∼ = commute, i.e., aOEb ⇔ ∃c(a ≡ c ∼ = b) ∼ iff ∃c(b ≡ c = a). Remark. Suppose Γ is given in terms of generators and relations, Γ = hx1 , x2 , . . . |w1 = 1, w2 = 1, . . . i, where x1 , x2 , . . . may be finite or infinite and similarly for w1 , w2 , . . . Then the space of homomorphisms of Γ into Aut(X, µ) can be identified, in either w or u, with the closed subspace of Aut(X, µ)N , where N is the cardinality of the set of generators, consisting of all {T1 , T2 , . . . } ∈ Aut(X, µ)N such that w1T1 ,T2 ... = w2T1 ,T2 ... = · · · = 1. Note that this subspace is conjugacy invariant and this identification preserves the conjugacy action of Aut(X, µ). In particular, the set of actions of FN , the free group on N generators, can be identified with Aut(X, µ)N . Finally, we denote by ERG(Γ, X, µ), WMIX(Γ, X, µ), MMIX(Γ, X, µ), MIX(Γ, X, µ), E0 RG(Γ, X, µ) the sets of ergodic, weak mixing, mild mixing, mixing, E0 -ergodic actions in A(Γ, X, µ). (These sets are of course empty if Γ is finite.) Also FR(Γ, X, µ), FRERG(Γ, X, µ) denote, resp., the sets of free, resp., free ergodic actions. Recall here that an action a(γ, x) = γ ·x of Γ on (X, µ) is ergodic if every invariant Borel set has measure 0 or 1. It is weak mixing if the product action γ · (x, y) = (γ · x, γ · y) on (X 2 , µ2 ) is ergodic. It is mild mixing if for any Borel set A ⊆ X with 0 < µ(A) < 1, limγ→∞ µ(A∆γ · A)) > 0. It is mixing if limγ→∞ µ(γ · A ∩ B) = µ(A)µ(B), for any two Borel sets A, B. It is E0 -ergodic or strongly ergodic if Ea is E0 -ergodic. Finally it is free if ∀γ 6= 1(γ · x 6= x, µ−a.e.(x)). There are many equivalent formulations of these notions that we will use in the sequel, for which we refer the reader to [BGo], [BR], [Gl2], [HK3], [Sc4]. (B) There is also another way to view the space of actions of Γ, studied in Glasner-King [GK] (see also Rudolph [Rud] for the case Γ = Z). We will take as our model space (X, µ) the space (T, λ), where T = unit circle and λ = (normalized) Lebesgue measure on T (the interval [0, 1] with Lebesgue
10. BASIC PROPERTIES
63
measure or the space 2N with the usual product measure would work as well). Consider also TΓ and the shift action s given by s(γ, p)(δ) = p(γ −1 δ). This is clearly a continuous action and we denote by SIM(Γ) the space of probability measures on TΓ which are shift-invariant. In the space of probability Borel measures on T endowed with the usual weak∗ -topology, SIM(Γ) is a compact, convex set. For any µ ∈ SIM(Γ) and γ ∈ Γ, let πγ : TΓ → T be the corresponding projection πγ (p) = p(γ), and consider (πγ )∗ µ. By shift-invariance this is independent of γ and we denote it as µ1 (= (π1 )∗ µ). It is called the marginal of µ. We denote by SIMν (Γ) the space of all µ ∈ SIM(Γ) with marginal ν. It is again a compact, convex set. For each a ∈ A(Γ, T, λ), x ∈ T, let ϕa (x) ∈ TΓ be the orbit map of x, i.e., ϕa (x)(γ) = a(γ −1 , x). Then ϕa : T → TΓ is a measurable injection (with ϕa (x)(1) = x), and ϕa (a(γ, x)) = s(γ, ϕa (x)), so (ϕa )∗ (λ) = µa is in SIM(Γ) with marginal λ. Moreover a on (T, λ) is isomorphic to s on (TΓ , µa ). Glasner-King show that Φλ (a) = µa is a homeomorphism of (A(Γ, T, λ), w) with a dense Gδ subset of SIMλ (Γ). Similarly, if µ is any non-atomic probability measure on T with support T, there is a corresponding homeomorphism Φµ (defined exactly as before) of A(Γ, T, µ) with a dense Gδ subset of SIMµ (Γ). Let H be the Polish group of orientation preserving homeomorphisms of T that fix 1. Then h ∈ H 7→ h∗ λ is a bijection of H with the set of all non-atomic, full support probability measures on T. Moreover, if for h ∈ H we let hΓ : TΓ → TΓ be defined by hΓ (p)(γ) = h(p(γ)), then the following diagram commutes Φµ
A(Γ, T, µ) −→ SIMµ (Γ) ¯ ↑h ↑ hΓ Φ
λ A(Γ, T, λ) −→ SIMλ (Γ)
¯ denotes the obvious homeomorphism of A(Γ, T, λ) with A(Γ, T, µ) where h ∼ ¯ ¯ induced by h, i.e., h(a)(γ, h(x)) = h(a(γ, x)). Note that h(a) = a. Now Glasner-King show that the map Φ : H × A(Γ, T, λ) → SIM(Γ), defined by ¯ Φ(h, a) = hΓ∗ (Φλ (a)) = Φh∗ λ (h(a)) is a homeomorphism of H × A(Γ, T, λ) with a dense Gδ subset of SIM(Γ). For further reference, we denote by ESIM(Γ) the set of ergodic measures in SIM(Γ). They are the extreme points of SIM(Γ) and they form a Gδ set in SIM(Γ). (C) Motivated by an analogous notion concerning unitary representations (see Appendix H), it is also useful to introduce a notion of weak containment for actions.
64
II. THE SPACE OF ACTIONS
Suppose a ∈ A(Γ, X, µ), b ∈ A(Γ, Y, ν). We say that a is weakly contained in b, in symbols a ≺ b, if for any A1 , . . . , An ∈ MALGµ , F ⊆ Γ finite, > 0, there are B1 , . . . , Bn ∈ MALGν such that |µ(γ a (Ai ) ∩ Aj ) − ν(γ b (Bi ) ∩ Bj )| < , ∀γ ∈ F, i, j ≤ n. It is clear that we can restrict here the A1 , . . . , An to belong to any countable dense subalgebra of MALGµ and also form a partition of X. We also say that a, b are weakly equivalent, in symbols a ∼ b, if a ≺ b and b ≺ a. We will now provide an equivalent characterization of weak containment (analogous to that in H.2). Proposition 10.1. Let a ∈ A(Γ, X, µ), b ∈ A(Γ, Y, ν). Then a ≺ b ⇔ a ∈ {c ∈ A(Γ, X, µ) : c ∼ = b}, where closure is in the weak topology. Proof. ⇐: Fix A1 , . . . , An ∈ MALGµ , F ⊆ Γ finite, > 0. Then there is c ∼ = b such that ∀γ ∈ F ∀i, j ≤ n(|µ(γ a (Ai ) ∩ Aj ) − µ(γ c (Ai ) ∩ Aj )| < ). If ϕ : (X, µ) → (Y, ν) is a Borel isomorphism that sends c to b, put ϕ(Ai ) = Bi . Then ϕ(γ c (Ai ) ∩ Aj ) = γ b (Bi ) ∩ Bj , so clearly ∀γ ∈ F ∀i, j ≤ n(|µ(γ a (Ai ) ∩ Aj ) − ν(γ b (Bi ) ∩ Bj )| < ). ⇒: By Section 1, (B), basic open nbhds of a in (A(Γ, X, µ), w) have the form V = {c : ∀γ ∈ F ∀i, j ≤ n(|µ(γ a (Ai ) ∩ Aj ) − µ(γ c (Ai ) ∩ Aj )| < )}, where A1 , . . . , An ∈ MALGµ form a partition of X, > 0 and F ⊆ Γ is finite containing 1 ∈ Γ. We need to find c ∼ = b in any such nbhd. Since a ≺ b, for any δ > 0 we can find B1 , . . . , Bn in MALGν such that ∀γ ∈ F ∀i, j ≤ n(|µ(γ a (Ai ) ∩ Aj ) − ν(γ b (Bi ) ∩ Bj )| < δ). Take δ <
. 20n3
Since 1 ∈ F , we have |µ(Ai ∩ Aj ) − ν(Bi ∩ Bj )| < δ, ∀i, j ≤ n,
P so ν(Bi ∩Bj ) < δ, ifSi 6= j ≤ n,T and |µ(Ai )−ν(Bi )| < δ, so |1− ni=1 ν(Bi )| < nδ. Let Bi0 = Bi \ j
10. BASIC PROPERTIES
65
αi ≤ µ(Ai ), and Ci ⊇ Ai if αi > µ(Ai ). Indeed, by renumbering, we can assume that A1 , . . . , Am are such that αi ≤ µ(Ai ), ∀i ≤ m, and Am+1 , . . . , An are such that µ(Aj ) < αj , ∀m + 1 ≤ j ≤ n. Let then Ci ⊆ Ai , P i ≤ m, be such that µ(Ci ) = αi . Let λi = µ(Ai ) − αi ≥ 0. Then clearly m i=1 λi = P n (α − µ(A )), so we can find for each m + 1 ≤ j ≤ n, C ⊇ Aj with j j j j=m+1 S m Cj \ Aj ⊆ j=1 (Ai \ Ci ) and µ(Cj ) = αj . 2 ¯i ), ∀i ≤ n, and So fix a partition C1 , . . . , Cn of X with µ(Ci ) = ν(B 3 µ(Ai ∆Ci ) < 5n δ. Let then ϕ : (Y, ν) → (X, µ) be a Borel isomorphism with ¯i ) = Ci , ∀i ≤ n. Let ϕ send b to c. Then for γ ∈ F, µ(γ c (Ci ) ∩ Cj ) = ϕ(B ¯i ) ∩ B ¯j ), and µ((γ c (Ci ) ∩ Cj )∆(γ c (Ai ) ∩ Aj )) < 10n3 δ, therefore ν(γ b (B a ¯i ) ∩ B ¯j )| + |µ(γ (Ai ) ∩ Aj ) − µ(γ c (Ai ) ∩ Aj )| ≤ |µ(γ a (Ai ) ∩ Aj ) − ν(γ b (B b c c c ¯ ¯ |ν(γ (Bi ) ∩ Bj ) − µ(γ (Ci ) ∩ Cj )| + |µ(γ (Ci ) ∩ Cj ) − µ(γ (Ai ) ∩ Aj )| < 5n3 δ + 10n3 δ < 20n3 δ < , so c ∈ V . 2 Thus if a, b ∈ A(Γ, X, µ), a ≺ b ⇔ a in the closure of the conjugacy class of b. It follows that ≺ on A(Γ, X, µ) is a special case of a familiar partial preordering, present in any continuous action. If a topological group G acts continuously on a topological space P , we have the following partial preordering on P : x 4 y ⇔ x ∈ G · y ⇔ G · x ⊆ G · y. The equivalence relation induced by 4, i.e., x ≈ y ⇔ x 4 y & y 4 x, is sometimes called the topological ergodic decomposition of this action. It is clear from the definition that each section 4y = {x : x 4 y} is closed and, if P is separable metrizable, then 4 is Gδ . For the particular case G = (Aut(X, µ), w), P = (A(Γ, X, µ), w) and the conjugacy action, we then obtain ≺. So we have the following fact. Corollary 10.3. The relation ≺ on A(Γ, X, µ) in Gδ and each section ≺b = {a ∈ A(Γ, X, µ) : a ≺ b} is closed in the weak topology. A basic property of weak containment, that will be useful later on, is stated in the next result. Let ai ∈ A(Γ, Xi , µi ), i ∈ I, Qwhere I is a countable index set, and ρ a probability Borel measure on i XiQwhose marginal on Xi is µi (i.e., Q the Xi -projection of ρ is µi ). Denote by i ai the product action of Γ on i Xi : γ
Q
i
ai
({xi }) = {γ ai (xi )}.
(Note that each A(Γ, Xi , µi ) is really a set of equivalence classes of Borel actions of Γ on Xi , where two actions are equivalent if they agree µi -a.e. Observe that if ai , a0i are equivalent with respect to Q Q now Qµi , for each i, then 0 are equivalent with respect to ρ.) If ρ is a , a then i i i i i ai -invariant, Q Q we say that ρ is a joining of {ai }. Q In this case i ai ∈ A(Γ, i Xi , ρ). To emphasize that we are looking at i ai with respect to ρ, we also write
66
II. THE SPACE OF ACTIONS
Q Q ( i ai )ρ . A special case is when ρ = i µi , in which case we simply write Q i ai . We note that the map Y Y Y Y {ai }i ∈ A(Γ, Xi , µi ) 7→ ai ∈ A(Γ, Xi , µi ) i
i
i
i
is continuous in the weak topology and, if I is finite, also continuous in the uniform topology. Proposition 10.4. Q Let ai ∈ A(Γ, Xi , µi ), i ∈ I, and ρ a joining of {ai }, so Q that ( i ai )ρ ∈ A(Γ, i Xi , ρ). Then Y ai ≺ ( ai )ρ , ∀i. i
In particular, ai ≺
Y
ai , ∀i.
i
Proof. Fix i ∈ I. Given A ∈ MALGµi , let Aˆ = π −1 (A) ∈ MALGρ , Q ˆ where π : i Xi → Xi is the projection. Clearly ρ(A) = µi (A) and Q ai (A). So given A , . . . , A ∈ MALG , if B = A ˆ = γ\ ˆj , we have γ ( i ai )ρ (A) 1 n µi j Q Q ( a ) a µi (γ i (Ak ) ∩ A` ) = ρ(γ i i ρ (Bk ) ∩ B` ), which shows that ai ≺ ( i ai )ρ . 2 Note also that ai ≺ bi , ∀i ⇒
Y
ai ≺
Y
i
bi .
i
Every a ∈ A(Γ, X, µ) gives rise to a unitary representation on L2 (X, µ), denoted by κa , called the Koopman representation associated to a. This is defined by κa (γ)(f )(x) = f ((γ −1 )a (x)). The map a 7→ κa ∈ Rep(Γ, L2 (X, µ)) is a homeomorphism of (A(Γ, X, µ), w) with the closed subspace of Rep(Γ, L2 (X, µ)) consisting of all representations π ∈ Rep(Γ, L2 (X, µ)) such that π(γ) ∈ Aut(X, µ) ⊆ U (L2 (X, µ)), ∀γ ∈ Γ (see Appendix H for a discussion of the space Rep(Γ, H) of unitary representations of a group Γ on a Hilbert space H). Let also κa0 be the restriction of κa to L20 (X, µ). Again a 7→ κa0 ∈ Rep(Γ, L20 (X, µ)) is a homeomorphism of (A(Γ, X, µ), w) with a closed subspace of Rep(Γ, L20 (X, µ)). Clearly if a, b are isomorphic, κa (resp., κa0 ), κb (resp., κb0 ) are isomorphic. We call a, b unitarily equivalent (or spectrally equivalent) if κa ∼ = κb . This is also a b ∼ equivalent to κ0 = κ0 . Note that a is ergodic (resp., weak mixing, mild mixing, mixing) iff κa0 is ergodic (resp., weak mixing, mild mixing, mixing) in the sense of Appendix H, (D). Recall also the notion of weak containment π ≺ ρ of unitary representations (see Appendix H). We have the following fact. Proposition 10.5. a ≺ b ⇒ κa ≺ κb , κa0 ≺ κb0 .
10. BASIC PROPERTIES
67
Proof. It is enough to check that a ≺ b ⇒ κa0 ≺ κb0 . We can assume that a, b ∈ A(Γ, X, µ) and view A(Γ, X, µ) as a closed subspace of Rep(Γ, L20 (X, µ)). If a ≺ b, then a is in the closure of the conjugacy class of b, i.e., in {T bT −1 : T ∈ Aut(X, µ)}, where closure is in the weak topology. Thus, since we identify a (resp., b) with κa0 (resp., κb0 ) ∈ Rep(Γ, L20 (X, µ)), we have −1
κa0 ∈ {UT0 κb0 (UT0 )
: T ∈ Aut(X, µ)} ⊆ {U κb0 U −1 : U ∈ U (L20 (X, µ))} = {σ ∈ Rep(Γ, L20 (X, µ)) : σ ∼ = κb0 }.
Then (see the proof of H.2) κa0 ≺ κb0 .
2
Remark. In fact this argument shows that a ≺ b ⇒ κa ≺Z κb , κa0 ≺Z κb0 , where ≺Z is the stronger notion of weak containment in the sense of Zimmer; see the Remark after Proposition H.2. We denote by iΓ ∈ A(Γ, X, µ) the trivial action of Γ on (X, µ) : γ · x = x. For the next proposition, recall the result of Jones-Schmidt [JS] that an ergodic action a ∈ A(Γ, X, µ) is not E0 -ergodic iff it admits non-trivial almost invariant sets. Here a ∈ A(Γ, X, µ) admits non-trivial almost invariant sets iff ∃δ > 0∀F finite ⊆ Γ∀ > 0∃A ∈ MALGµ [δ < µ(A) < 1 − δ and ∀γ ∈ F (µ(γ a (A)∆A) < )] iff ∃δ > 0∃{An } ⊆ MALGµ [δ < µ(An ) < 1 − δ, and ∀γ ∈ F (µ(γ a (An )∆An ) → 0)]. (For a survey of these notions and a proof of this result see also Hjorth-Kechris [HK3], Appendix A.) Moreover, if a is ergodic but not E0 -ergodic, we can find such almost invariant sets of any given measure 0 < c < 1. Proposition 10.6. Let a ∈ A(Γ, X, µ). If iΓ ≺ a, then a admits nontrivial almost invariant sets. Moreover if a is ergodic, then iΓ ≺ a ⇔ a is not E0 -ergodic. Proof. First assume iΓ ≺ a. Fix F ⊆ Γ finite containing 1 and > 0. Let also A ∈ MALGµ be such that µ(A) = 1/2. Put A1 = A, A2 = X \ A. Then for any δ > 0, there are B1 , B2 ∈ MALGµ such that µ(B1 ∩ B2 ) < δ, 1 1 | − µ(B1 )|, | − µ(B2 )| < δ, 2 2 µ(B1 ∩ γ a (B2 )), µ(B2 ∩ γ a (B1 )) < δ, ∀γ ∈ F . If we choose δ small enough, then clearly µ(γ a (B1 )∆B1 ) < , ∀γ ∈ F . So a has non-trivial almost invariant sets. Conversely, if a is ergodic but not E0 -ergodic, then the proof of the above result of Jones-Schmidt (see again Hjorth-Kechris [HK3], A2.2) shows that for any partition A1 , . . . , An of X, with say µ(Ai ) = αi , and every
68
II. THE SPACE OF ACTIONS
> 0, F ⊆ Γ finite, symmetric containing 1, there is a partition B1 , . . . , Bn of X with µ(Bi ) = αi and µ(γ a (Bi )∆Bi ) < for γ ∈ F . Thus for i, j ≤ n, |µ(Ai ∩ Aj ) − µ(γ a (Bi ) ∩ Bj )| < , 2
so iΓ ≺ a.
If 1Γ is the trivial 1-dimensional representation of Γ, then 1Γ ≺ κa0 ⇔ κa0 admits non-0 almost invariant vectors. (where a representation π admits non-0 almost invariant vectors if ∀F ⊆ Γ finite ∀ > 0∃v (kπ(γ)(v) − vk < kvk). If iΓ ≺ a, then, as κi0Γ = ∞ · 1Γ , we have 1Γ ≺ κa0 . However, 1Γ ≺ κa0 does not imply iΓ ≺ a, since there are examples of ergodic actions of Γ = F2 with 1Γ ≺ κa0 but iΓ 6≺ a (see Schmidt [Sc3], 2.7 and Hjorth-Kechris [HK3], A3.2). Thus, in general, κa0 ≺ κb0 does not imply a ≺ b. (D) In Glasner-King [GK], p. 234, the authors asked for what groups Γ there is a dense conjugacy class in (A(Γ, X, µ), w). They note that Rudolph and Weiss pointed out that this holds for all amenable Γ and, in particular, Glasner-King asked if this is also the case for any free group. It turns out an affirmative answer for the free group case is an immediate consequence of Kechris-Rosendal [KR], 6.5, but in fact it is always true, as shown, independently, by Hjorth (unpublished) and Glasner-Thouvenot-Weiss [GTW]. Theorem 10.7 (Glasner-Thouvenot-Weiss, Hjorth). For each countable group Γ, the space A(Γ, X, µ) has a weakly dense conjugacy class. Proof. Fix a countable Q dense set {an } in (A(Γ, X, µ), w) and consider the product action a = n an of Γ on (X N , µN ) : a(γ, {xn }) = {an (γ, xn )}. Then, by 10.4, an ≺ a, ∀n, so as {an } is dense in A(Γ, X, µ), b ≺ a, for every b ∈ A(Γ, X, µ), i.e., an isomorphic copy of a in A(Γ, X, µ) has weakly dense conjugacy class. 2 Also note the following result that can be proved by a similar method. Theorem 10.8 (Glasner-King). Let Γ be a countable group. The set FR(Γ, X, µ) of free actions in A(Γ, X, µ) is dense Gδ in the weak topology. Proof. First note that FR(Γ, X, µ) is Gδ . We have a ∈ FR(Γ, X, µ) ⇔ ∀γ 6= 1(δu (γ a , 1) = 1) and open balls in δu are Fσ in the weak topology. To check that FR(Γ, X, µ) is dense, note that Γ admits some free action, e.g., the shift on 2Γ if Γ is infinite or the translation action γ · (δ · x) = (γδ, x) on Γ×X if Γ is finite. Thus, if in the proof Q of 10.7 we include in the sequence {an } a free action, it follows that a = n an is free and its isomorphic copies are dense in A(Γ, X, µ). 2 (E) We next consider factors and extensions of actions. Given two actions a ∈ A(Γ, X, µ), b ∈ A(Γ, Y, ν), a homomorphism of a to b is a Borel map π : X → Y such that π∗ µ = ν and π(γ a (x)) = γ b (π(x)), µ-a.e. (x), ∀γ ∈ Γ.
10. BASIC PROPERTIES
69
If such a homomorphism exists, then b is a factor of a and a an extension of b. We denote this by b v a. We note that for π as above the map π ∗ (A) = π −1 (A) is an embedding of MALGν to a Γ-invariant non-atomic σ-subalgebra of MALGµ . Conversely, every Γ-invariant non-atomic σ-subalgebra A ⊆ MALGµ gives rise to a homomorphism π from a to some b, so that π ∗ is an isomorphism of MALGν with A which preserves the action of Γ (see Mackey [M] or Kechris [Kec1], 3.47). Thus there is a canonical bijection between factors of a and Γ-invariant non-atomic σ-subalgebras of MALGµ . Simple examples of factors come from joinings. If ρ is a joining of {ai }, Q clearly ai v ( i ai )ρ , ∀i. There is a canonical way to produce extensions of a given action. Let b ∈ A(Γ, Y, ν) be a given action, let (Z, σ) be a standard Borel space with a probability measure σ, not necessarily non-atomic, and let β : Γ × Y → Aut(Z, σ) be a Borel cocycle of the action b into Aut(Z, σ), i.e., a Borel map satisfying β(γ1 γ2 , y) = β(γ1 , γ2b (y))β(γ2 , y), a.e.(y), ∀γ1 , γ2 ∈ Γ. Let X = Y × Z, µ = ν × σ and define the skew product action a ∈ A(Γ, X, µ) of b with β, denoted by b ×β (Z, σ), by γ a (y, z) = (γ b (y), β(γ, y)(z)). Clearly b v a. Conversely a theorem of Rokhlin (see, e.g., Glasner [Gl2], 3.18) asserts that if b v a and a is ergodic, then a ∼ = b ×β (Z, σ) for some (Z, σ), β as above. We note that the argument in the proof of 10.4 clearly shows that a v b ⇒ a ≺ b. Moreover, it is easy to check that a v b ⇒ κa ≤ κb , κa0 ≤ κb0 . We call a, b weakly isomorphic, in symbols a∼ =w b if a v b and b v a. Thus we have ∼b⇒a∼ a= =w b ⇒ a ∼ b, a∼ =w b ⇒ κa ∼ = κb . For Γ = Z and a, b free and ergodic, it is well-known that none of these implications reverses. For examples of a, b with a ∼ =w b but a 6∼ = b see, e.g., Glasner [Gl2], 7.16. Since a ∼ b for all such a, b, clearly a ∼ b 6⇒ a ∼ =w b. Finally, weakly isomorphic actions of Z have equal entropies and all Bernoulli shifts on 2Z are spectrally equivalent, so the last implication does not reverse.
70
II. THE SPACE OF ACTIONS
However, I do not know if for every infinite, countable group Γ, there are a, b ∈ FRERG(Γ, X, µ) with a ∼ b but a ∼ 6 b, and also whether there are = a b ∼ ∼ a, b ∈ FRERG(Γ, X, µ) with κ = κ but a = 6 w b. Remark. A natural question, raised by Tsankov, is whether for a given infinite countable group Γ there is a∞ ∈ A(Γ, X, µ) such that for any a ∈ FRERG(Γ, X, µ), a v a∞ . The answer is however negative. If such a∞ existed, then κa0 ≤ κa0∞ for all a ∈ FRERG(Γ, X, µ) and thus, by E.1 and the paragraph following it, if π is an irreducible infinite-dimensional representation of Γ, there is a ∈ FRERG(Γ, X, µ) with π ≤ κa0 ≤ κa0∞ . If Γ is not abelian by-finite, then (see, e.g., Thoma [Th] or Hjorth [Hj3]) Γ has uncountably many non-isomorphic infinite-dimensional irreducible representations, which immediately gives a contradiction. Consider now the case where Γ is abelian-by-finite and fix ∆ Γ, [Γ : ∆] < ∞, ∆ abelian. We use below the notation and facts of Appendix G. For any a ∈ FRERG(∆, X, µ), IndΓ∆ (a) ∈ FRERG(∆, X × T, µ × νT ). Thus Γ IndΓ∆ (a) v a∞ and so κInd∆ (a) ∼ = IndΓ∆ (κa ) ≤ κa∞ and therefore IndΓ∆ (κa )|∆ ≤ κa∞ |∆. Let π = κa∞ |∆. Using now Appendix F, it follows that the maximal spectral type of each IndΓ∆ (κa )|∆ is the maximal spectral type of π, say ρ, and thus, by the last paragraph of Appendix G, the maximal spectral type of κa is ρ. Thus we have derived that all maximal spectral types of (the Koopman representations of) free, ergodic measure preserving actions of ∆ ˆ But this is clearly false. are absolutely continuous to a fixed measure ρ on ∆. ˆ For example, given any non-atomic symmetric probability measure ν on ∆ −1 (symmetric means invariant under x 7→ x ), ϕ = νˆ is real positive-definite on ∆ (see Appendix F) and if we consider the shift action aν of ∆ on (R∆ , µϕ ) (see Appendix D), then the maximal spectral type of aν is ν. Since, whenever ν is non-atomic, aν is free, ergodic (see, e.g., Cornfeld-FominSinai [CFS], 14.2.1), it follows that every non-atomic symmetric measure on ˆ is absolutely continuous to ρ, which is absurd, since ∆ ˆ is uncountable. ∆ (F) We can also endow A(Γ, X, µ) with a concept of “convex P combination”. Let a1 , . . . , an ∈ A(Γ, X, µ) and 0P ≤ λi ≤ 1, 1 ≤ i ≤ n, ni=1 λi = 1. Then we define, up to isomorphism, b = ni=1 λi ai ∈ A(Γ, X, µ) as follows: ¯ =X ¯ t ··· t X ¯ n , where each X ¯ i is a copy of X, say via a copying Let X P ¯ as µ map x 7→ x ¯i . Let µ ¯i be the copy of µ on XiS . Define µ ¯ on X ¯ = ni=1 λi µ ¯i . P Let a ¯i be the copy of ai on Xi and let a ¯ = ni=1 a ¯i . Then ni=1 λi ai is defined to be any b ∈ A(Γ, X, µ) such that b ∼ ¯. (One can also define integrals of =a actions over a probability measure but we will not discuss this here.) ¯1, . . . , X ¯ n as above give the ergodic When each ai is ergodic, then X P decomposition of a ¯. Thus the set of all convex combinations ni=1 λi ai for ergodic ai gives, up to isomorphism, all the actions a ∈ A(Γ, X, µ) which have a finite ergodic decomposition. Denote the set of such actions by FED(Γ, X, µ).
10. BASIC PROPERTIES
71
Proposition 10.9. The set FED(Γ, X, µ) of all actions in A(Γ, X, µ) that have a finite ergodic decomposition is dense in (A(Γ, X, µ), w). Proof. We can assume that X = 2N . Fix an action a ∈ A(Γ, X, µ) in order to show that a is in the weak closure of FED(Γ, X, µ). Let ϕa : X → X Γ be the usual orbit map ϕa (x)(γ) = a(γ −1 , x). Then ϕa is a Borel embedding of the action a into the shift action s of Γ on X Γ , which is clearly continuous. Let κ = (ϕa )∗ µ. Then s ∈ A(Γ, X Γ , κ) and s ∼ = a. Thus, by replacing a by s, we can assume without loss of generality that X = 2N (as X Γ is homeomorphic to X) and a is continuous. Consider the compact convex set of probability measures on X which are a-invariant, with the weak∗ topology. The extreme points of that set are the probability measures Pm that are a-invariant, ergodic, so by Krein-Milman the set of measures i=1 αi µi , P α = 1 and each µ is a-invariant, ergodic, is dense where 0 ≤ αi ≤ 1, m i i=1 i in the set of a-invariant probability measures. Since the Boolean algebra of clopen sets in 2N is dense in the measure algebra MALGµ , a basic open nbhd of a in A(Γ, X, µ) can be taken to be of the form V = {b : ∀γ ∈ F ∀j, k ≤ n|µ(γ a (Aj ) ∩ Ak ) − µ(γ b (Aj ) ∩ Ak )| < }, where A1 , . . . , An is a clopen partition of 2N , F ⊆ Γ is finite containing 1 and > 0. We need to find Pb ∈ V ∩ FED(Γ, X, µ). Find αi , and distinct µi as above, such that if ν = m i=1 αi µi , then |µ(C) − ν(C)| < /3 for any C in the Boolean algebra generated by {γ a (Aj ) : γ ∈ F, 1 ≤ j ≤ n}. Fix a Borel partition Y1 ∪ · · · ∪ Ym = X, so that µi (Yi ) = 1 and each Yi is a-invariant. Then let Y¯i be a copy of Yi with copying map y 7→ y¯i , let µ ¯i be the copy of µi ¯ = Y¯1 t· · ·t Y¯m be the disjoint union of the Y¯i , and let µ ¯ be on Y¯i , let X ¯ on X P ¯ α µ ¯ (so that µ ¯ ( Y ) = α ). Let a ¯ be the copy of a|Y defined by µ ¯=Sm i i i i on i=1 i i m ∼ ¯ ¯ ¯i . Clearly a ¯ ∈ FED(Γ, Y , µ ¯). We will find a ¯ b ∈ V. Yi and let a ¯ = i=1 a Sm a¯i = S a ¯ (A ¯ γ (A A ∩ Y . Then γ ) = For A ⊆ X, let A¯ = m i j j ∩ Yi ) = i=1 Sm a Sm i=1a a ¯ ¯ j ) ∩ Yi = γ (Aj ). Clearly A1 , . . . , Am is a i=1 γ (Aj ∩ Yi ) = i=1 γ (A P P m m ¯ and µ partition of X ¯(A¯j ) = ¯i (Aj ∩ Yi ) = i=1 αi µ i=1 αi µi (Aj ∩ Yi ) = ν(Aj ). Similarly, µ ¯(γ a¯ (A¯j ) ∩ A¯k ) = µ ¯(γ a (Aj ) ∩ A¯k ) =µ ¯(γ a (Aj ) ∩ Ak ) =
m X
αi µ ¯i (γ a (Aj ) ∩ Ak ∩ Yi )
i=1
=
m X
αi µi (γ a (Aj ) ∩ Ak ∩ Yi )
i=1
= ν(γ a (Aj ) ∩ Ak ). So |¯ µ(A¯j ) − µ(Aj )| < /3, thus, by Lemma 10.2, there is a Borel partition B1 , . . . , Bm of X with µ(Bj ∆Aj ) < /3 and µ(Bj ) = µ ¯(A¯j ). Fix then an ¯ ¯ isomorphism ϕ : (X, µ ¯) → (X, µ) with ϕ(Aj ) = Bj and let the image of
72
II. THE SPACE OF ACTIONS
a ¯ under ϕ be b ∈ A(Γ, X, µ). Thus µ(γ b (Bj ) ∩ Bk ) = µ ¯(γ a¯ (A¯j ) ∩ A¯k ) = a a b ν(γ (Aj ) ∩ Ak ), so |µ(γ (Aj ) ∩ Ak ) − µ(γ (Bj ) ∩ Bk )| < 3 . We will check that b ∈ V . We have for γ ∈ F, j, k ≤ n, |µ(γ a (Aj ) ∩ Ak ) − µ(γ b (Aj ) ∩ Ak )| ≤ |µ(γ a (Aj ) ∩ Ak ) − µ(γ b (Bj ) ∩ Bk )| + |µ(γ b (Bj ) ∩ Bk ) − µ(γ b (Aj ) ∩ Ak )| 2 = . < + 3 3
2
Note also the following simple fact: ∀i ≤ n(ai ≺ bi ) ⇒
n X i=1
where
Pn
i=1 λi
λi ai ≺
n X
λ i bi ,
i=1
= 1, 0 ≤ λi ≤ 1.
(G) Next we discuss the relation between A(Γ, X, µ) and A(∆, X, µ) for certain pairs of groups Γ, ∆. First, we note that the space A(Γ1 ∗ Γ2 ∗ · · · ∗ Γn , X, µ) with the weak (resp., uniform) topology is canonically homeomorphic to A(Γ1 , X, µ)×· · ·× A(Γn , X, µ) with the product of the weak (resp., uniform) topologies and this homeomorphism preserves the conjugacy actions of Aut(X, µ) (where in the second case we of course look at the product action). Second, if Γ ≤ ∆, then there is a canonical continuous map π∆,Γ : A(∆, X, µ) → A(Γ, X, µ) defined by restriction: π∆,Γ (a) = a|∆. Again we can use here either the weak or the uniform topology. Moreover π∆,Γ clearly preserves the conjugacy action of Aut(X, µ). In case ∆ = Γ ∗ Γ0 , for some Γ0 , clearly this map is open and surjective, as in this case A(∆, X, µ) can be identified with the product A(Γ, X, µ) × A(Γ0 , X, µ) and π∆,Γ is simply the projection map to the first factor. Conversely, when Γ ≤ ∆, every action of Γ gives rise to an action of ∆ called the co-induced action. (This was brought to my attention by Adrian Ioana, who referred to Gaboriau [Ga2] in his paper Ioana [I1]. Sinelshchikov pointed out that this concept also appears in L¨ uck [Lu], and references therein, in another, non-measure theoretic context.) We will actually consider a somewhat more general situation. Suppose that a countable group ∆ acts on a countable set T , and let σ : ∆ × T → Aut(X, µ) be a cocycle of the action of ∆ on T with values in Aut(X, µ), i.e., a map satisfying the property σ(δ1 δ2 , t) = σ(δ1 , δ2 · t)σ(δ2 , t), for δ1 , δ2 ∈ ∆, t ∈ T . We define an action of ∆ on (Y, ν), where Y = X T , ν = µT (the product measure), by (δ · f )(t) = σ(δ −1 , t)−1 (f (δ −1 · t)).
10. BASIC PROPERTIES
73
It is straightforward to check that this is indeed a measure preserving action of ∆ on (Y, ν). We note that if τ : ∆ × T → Aut(X, µ) is cohomologous to σ, i.e., there is f : T → Aut(X, µ) such that τ (δ, t) = f (δ · t)σ(δ, t)f (t)−1 , then the action induced by τ is isomorphic to the action induced by σ via the map ϕ : Y → Y given by ϕ(p)(t) = f (t)(p(t)). If now Γ is another countable group and a ∈ A(Γ, X, µ), while ρ : ∆ × T → Γ is a cocycle of the action of ∆ on T with values in Γ, define the cocycle σ : ∆ × T → Aut(X, µ) by σ(δ, t) = ρ(δ, t)a . We call the action defined as above using this σ the co-induced action of a with respect to the action of ∆ on T and ρ. Consider now the case where Γ ≤ ∆ and let T be a transversal for the left cosets of Γ with 1 ∈ T . Let ∆ act on T by defining δ · t to be the unique element of T in δtΓ, and let ρ : ∆×T → Γ be the cocycle defined by ρ(δ, t) = (the unique γ ∈ Γ such that (δ · t)γ = δt). In that case we write CInd∆ Γ (a) for the co-induced action corresponding to this action and cocycle. Clearly ∼ b ⇒ CInd∆ (a) ∼ a= = CInd∆ (b). Γ
Γ
We note certain facts about this co-induced action (for proofs see Ioana [I1]): (i) a v CInd∆ Γ (a)|Γ. Indeed the map f 7→ f (1) shows that a is a factor of CInd∆ Γ (a)|Γ. (ii) a 7→ CInd∆ Γ (a) A(Γ, X, µ) → A(∆, Y, ν) is a homeomorphism of (A(Γ, X, µ), w) onto a (necessarily Gδ ) subspace of (A(∆, Y, ν), w). It is also a homeomorphism in the uniform topology, if [∆ : Γ] < ∞. (iii) Assume Γ is infinite. Then a is mixing (resp., mild mixing) iff ∆ CInd∆ Γ (a) is mixing (resp., mild mixing). If [∆ : Γ] = ∞, then CIndΓ (a) is weak mixing. If [∆ : Γ] < ∞, then CInd∆ Γ (a) is weak mixing iff a is weak mixing. In particular, if a is weak mixing, so is CInd∆ Γ (a). (iv) a is free iff CInd∆ (a) is free. Γ (v) If iΓ is the trivial action of Γ on X, then CInd∆ Γ (iΓ ) is the shift action of ∆ on X ∆/Γ . When Γ ∆, then for any γ ∈ Γ, t ∈ T, γ · t = t and ρ(γ, t) = t−1 γt, −1 so that CInd∆ Γ (a)|Γ is given by (γ · f )(t) = (t γt) · f (t). Thus if we define for each t ∈ TQthe action at ∈ A(Γ, X, µ) by at (γ, x) = a(t−1 Qγt, x), then ∆ T CInd∆ (a)|Γ = a . Moreover, if Γ ≤ Z(∆), CInd (a)|Γ = t Γ Γ t∈T t∈T a = a .
74
II. THE SPACE OF ACTIONS
In particular, this shows that a being ergodic does not always imply that CInd∆ Γ (a)|Γ is ergodic. Ioana has also shown that even a being weak mixing does not always imply that CInd∆ Γ (a)|Γ is ergodic. Since a v CInd∆ (a)|Γ, it follows that a ≺ CInd∆ Γ Γ (a)|Γ, so in particular the set of b ∈ A(Γ, X, µ) of the form b = c|Γ, for some c ∈ A(∆, X, µ) (i.e., the set of actions of Γ that can be extended to actions of ∆) is weakly dense in A(Γ, X, µ). In other words, the restriction map π∆,Γ defined above has dense range in the weak topology. Moreover, if [∆ : Γ] = ∞ the set of b ∈ A(Γ, X, µ) of the form b = c|Γ for some c ∈ WMIX(∆, X, µ) is also weakly dense in A(Γ, X, µ). (As we will see later, when Γ has property (T), WMIX(Γ, X, µ) is closed nowhere dense; see 12.8.) It follows from (ii), (iii) above and 2.7 that MMIX(Γ, X, µ) is (coanalytic) but not Borel, if Γ has an element of infinite order. In particular, mixing, mild mixing and weak mixing are all distinct for such groups (since MIX(Γ, X, µ), WMIX(Γ, X, µ) are Borel). It is unknown if this holds for every infinite group; see Section 11, (C). Remark. Let Γ ≤ ∆ be countable groups and denote by A(Γ ↑ ∆, X, µ) the set of actions in A(Γ, X, µ) that can be extended to an action of ∆. We have just seen that A(Γ ↑ ∆, X, µ) is dense in (A(Γ, X, µ), w). It is an interesting question to find out when the generic action of Γ extends to an action of ∆. For example, Ageev [A1] has shown that this is the case when Γ = Z and ∆ ≥ Γ is abelian. On the other hand one can see that there are cases in which the answer is negative. Let Γ = F4 with generators α, β, γ, δ. The generic action a of Γ has the following properties: It is free, so in particular αa , β a , γ a , δ a are distinct, and the group generated by αa , β a is dense in Aut(X, µ) (see Section 4, (D)). Let ϕ be the automorphism of Γ such that ϕ(α) = α, ϕ(β) = β, ϕ(γ) = δ, ϕ(δ) = γ. Let ∆ be the HNN extension of Γ corresponding to ϕ (i.e., the semidirect product of Z with Γ, where Z acts on Γ by n · γ = ϕn (γ)). Let t be the additional generator, so that tγt−1 = ϕ(γ), ∀γ ∈ Γ. If the generic action a ∈ A(Γ, X, µ) was extendible to an action b of ∆, let αb = αa = A, β b = β a = B, γ b = γ a = C, δ b = δ a = D, tb = T . Then T AT −1 = A, T BT −1 = B. so T commutes with every element of hA, Bi and thus every element of Aut(X, µ), so T = 1. But T CT −1 = D, i.e., C = D, a contradiction. We also recall from Appendix G that when Γ ≤ ∆ has finite index, we can also define for each a ∈ A(Γ, X, µ) the induced action Ind∆ Γ (a), defined as follows: Let T, ρ be as before (for the pair Γ, ∆), νT be the normalized counting measure on T . Then Ind∆ Γ (a) ∈ A(Γ, X × T, µ × νT ) is defined by δ · (x, t) = (ρ(δ, t) · x, δ · t). If a is free (resp., ergodic), so is Ind∆ Γ (a). Finally, let Γ be a factor of ∆. The clearly there is a topological embedding (in either the weak or the uniform topology) ρΓ,∆ : A(Γ, X, µ) → A(∆, X, µ),
10. BASIC PROPERTIES
75
which preserves conjugacy and such that the equivalence relation generated by a is the same as that generated by ρΓ,∆ (a). (H) We conclude this section with some results and questions concerning the theory of costs. General references for this theory are Gaboriau [Ga1] and Kechris-Miller [KM]. We use the terminology and notation of [KM]. Let Γ be an infinite countable group. For each a ∈ A(Γ, X, µ), let the cost of a be defined by C(a) = Cµ (Ea ), i.e., the cost of a is the cost of the associated equivalence relation. Thus aOEb ⇒ C(a) = C(b). Also 0 ≤ C(a) ≤ ∞. The cost of Γ is defined by C(Γ) = inf{C(a) : a ∈ FR(Γ, X, µ)}. Moreover, it turns out that C(Γ) = inf{C(a) : a ∈ FRERG(Γ, X, µ)} = min{C(a) : a ∈ FRERG(Γ, X, µ)}, see Kechris-Miller [KM], 29.1 and 29.2. We have 1 ≤ C(Γ) ≤ ∞. A group Γ is said to have fixed price if C|FR(Γ, X, µ) is constant. It is unknown whether every infinite countable group has fixed price. However, the calculations in 18.1 of Kechris-Miller [KM] show that for each r ∈ R the set Ar = {a ∈ A(Γ, X, µ) : C(a) < r} is analytic in (A(Γ, X, µ), w) and thus, in particular, C is a Baire measurable function on this space. The following is then immediate from 10.7, 10.8 and the fact that C is conjugacy invariant (see Kechris [Kec2], 8.46). Proposition 10.10. Let Γ be an infinite countable group. Then there is a dense Gδ set in (FR(Γ, X, µ), w) on which the cost is constant. In particular, if C is continuous on (FR(Γ, X, µ), w), then Γ has fixed price. Denote this value C ∗ (Γ). Clearly C(Γ) ≤ C ∗ (Γ). Problem 10.11. Is C(Γ) = C ∗ (Γ)? Remark. Using 12.1 and 13.1 below, we also see that there is a dense Gδ set in (FRERG(Γ, X, µ), w) on which the cost is constant. Call this fixed value C ∗∗ (Γ). Clearly C(Γ) ≤ C ∗∗ (Γ). If Γ does not have property (T) then 12.2, ii) shows that FRERG(Γ, X, µ) is dense Gδ in (FR(Γ, X, µ), w), so C ∗ (Γ) = C ∗∗ (Γ), but if Γ has property (T), 12.2, i) shows that the set FRERG(Γ, X, µ) is closed nowhere dense in (FR(Γ, X, µ), w) and it is not clear how the quantities C ∗ (Γ), C ∗∗ (Γ) are related in this case. In connection with 10.11, let us note the following fact. Proposition 10.12. Let Γ be an infinite countable group. Then MINCOST(Γ, X, µ) = {a ∈ FR(Γ, X, µ) : C(a) = C(Γ))
76
II. THE SPACE OF ACTIONS
is dense in (FR(Γ, X, µ), w). Proof. It has been shown by Gaboriau [Ga1] (see also Kechris-Miller Q ∼ [KM], 29.1) that if b = n an , then C(b) ≤ C(an ), ∀n. Let then {an } ⊆ FR(Γ,Q X, µ) be dense and such that C(Γ) = inf{C(an ) : n ∈ N}. By 10.4, if b ∼ = n an , then the conjugacy class of b is dense in (FR(Γ, X, µ), w) and C(b) = C(Γ). 2 Clearly if MINCOST(Γ, X, µ) is Gδ we have an affirmative answer to 10.11. We next show that 10.11 has indeed a positive answer if Γ is finitely generated. The main fact is the following result which gives an estimate for the complexity of the cost function. Theorem 10.13 (Kechris). Let Γ be an infinite, finitely generated group. Then the cost function C is upper semicontinuous on (FR(Γ, X, µ), w), i.e., for each r ∈ R, Ar ∩ FR(Γ, X, µ) = {a ∈ FR(Γ, X, µ) : C(a) < r} is open in (FR(Γ, X, µ), w). Proof. We will use the arguments in 18.1 and 18.3 in Kechris-Miller [KM] together with a modification due to Melleray (see Corrections and Updates for Kechris-Miller [KM], www.math.caltech.edu/people/kechris.html). Let Γ = hγ1 , . . . , γN i and let Γ = {gi }i∈N be an enumeration of Γ. Fix a countable Boolean algebra {A` } ⊆ MALGµ , which is dense in MALGµ , let n 7→ ((n)0 , (n)1 ) be a bijection of N with N × N and for a ∈ A(Γ, X, µ), let a |A a a {θna } = {g(n) (n)1 }. For S ⊆ N, put ΘS = {θn }n∈S . Then one can see that 0 C(a) < r ⇔ ∃ finite T ⊆ N∃M [Cµ (ΘaT ) + Cµ ({γia |D(γia , ΘaM,T )}i≤N ) < r], where ±1 ΘaM,T = {θ1±1 · · · θM : θi ∈ ΘaT ∨ θi = id, ∀i ≤ M }
and D(γia , ΘaM,T ) = {x : γia (x) 6= θ(x), ∀θ ∈ ΘaM,T }. Indeed, if such T, M exist, then, as ΘaT t {γia |D(γia , ΘaM,T )}i≤N is obviously an L-graphing of Ea and has cost < r, clearly C(a) < r. Conversely, assume C(a) < r. Then, by the first paragraph of the proof of 18.1 in [KM], there is an L-graphing of Ea of the form ΘaS , S ⊆ N, such that Cµ (ΘaS ) = α < r. Let < r − α. Next notice that, since ΘaS is an L-graphing of Ea , for each i ≤ N , lim µ(D0 (γia , ΘaS∩{0,...,n} ) = 0,
n→∞
where D0 (θ, Θ) = {x : (x, θ(x)) 6∈ RΘ }, with RΘ the equivalence relation generated by Θ. So choose L large enough, so that if T = S ∩ {0, . . T . , L}, µ(D0 (γia , ΘaT )) < N , ∀i ≤ N . Furthermore, since D0 (γia , ΘaT ) = M D(γia , ΘaM,T ), we can find M large enough
10. BASIC PROPERTIES
77
so that µ(D(γia , ΘaM,T )) < N , ∀i ≤ N . Then Cµ (ΘaT ) ≤ α < r − and Cµ ({γia |D(γia , ΘaM,T )}i≤N ) < N · N = , so we are done. Now X Cµ (ΘaT ) = µ(A(n)1 ) = αT n∈T
is independent of a. So it is enough to show that for each s ∈ R, finite T ⊆ N, and M ∈ N, the set {a ∈ FR(Γ, X, µ) : Cµ ({γia |D(γia , ΘaM,T )}i≤N ) < s} is open in (FR(Γ, X, µ), w). Since Cµ ({γia |D(γia , ΘaM,T )}i≤N ) =
X
µ(D(γia , ΘaM,T )),
i≤N
and µ : MALGµ → [0, 1] is continuous, it is enough to show that for each fixed i ≤ N , finite T ⊆ N, M ∈ N, a 7→ D(γia , ΘaM,T ) FR(Γ, X, µ) → MALGµ is continuous in the weak topology. Inspecting the definition of D(γia , ΘaM,T ), we see that there is a finite sequence δ1 , . . . , δk ∈ Γ, a finite sequence n1 , . . . , n` ∈ N, and a Boolean formula ϕ(x1 , . . . , xk , y11 , . . . , y1` , y21 , . . . , y2` , . . . , yk1 , . . . , yk` ), all independent of a, such that D(γia , ΘaM,T ) = ϕ(supp(δ1a ), . . . , supp(δka ), δ1a (An1 ), . . . , δ1a (An` ), δ2a (An1 ), . . . , δ2a (An` ), ... δka (An1 ), . . . , δka (An` )). Recall that the Boolean operations are continuous in MALGµ . Since a 7→ δ a (A) is continuous from (A(Γ, X, µ), w) into MALGµ , for each δ ∈ Γ, A ∈ MALGµ , and since supp(δ a ) is either ∅ or X on FR(Γ, X, µ), depending on whether δ = 1 or δ 6= 1 (and this is the only place where we use freeness of the actions), it follows that a 7→ D(γia , ΘaM,T ) is continuous and the proof is complete. 2 For a, b ∈ FR(Γ, X, µ), put a b ⇔ a ∈ {c ∈ FR(Γ, X, µ) : cOEb}, where closure is in the weak topology. Thus a ≺ b ⇒ a b.
78
II. THE SPACE OF ACTIONS
Corollary 10.14. Let Γ be an infinite, finitely generated group. For a, b ∈ FR(Γ, X, µ), a b ⇒ C(a) ≥ C(b). In particular, if a ∈ FR(Γ, X, µ) has weakly dense orbit equivalence class, restricted to free actions, e.g., if a has weakly dense conjugacy class, then C(a) = C(Γ). Therefore, C ∗ (Γ) = C(Γ). Moreover, MINCOST(Γ, X, µ) is dense Gδ in the weak topology and it is equal to the set of continuity points of C|FR(Γ, X, µ). Ab´ert and Weiss (unpublished) have recently shown that for any infinite group Γ, the shift action sΓ of Γ on (X, µ), where X = 2Γ and µ is the usual product measure, is minimum in ≺ among free actions of Γ. This implies by the above that for infinite, finitely generated groups Γ, the action sΓ has the maximum cost among free actions of Γ. Remark. The cost function is not upper semicontinuous in the space (A(Γ, X, µ), w), for any infinite Γ. Otherwise, if a ∈ FR(Γ, X, µ) has weakly dense conjugacy class, then 1 ≤ C(a) ≤ C(b) for any b ∈ A(Γ, X, µ), which gives a contradiction by taking b = iΓ (the trivial action) for which C(b) = 0. On the other hand, since the map T 7→ supp(T ) from Aut(X, µ) to MALGµ is continuous in the uniform topology u (see Section 1), it follows that a 7→ supp(δ a ) is continuous from (A(Γ, X, µ), u) to MALGµ , ∀δ ∈ Γ. Then, by the argument in the proof of 10.13, the cost function C is upper semicontinuous in (A(Γ, X, µ), u) for any infinite, finitely generated group. Remark. Tsankov pointed out the following fact, relating extensions of actions and cost. If a ∈ FR(Γ, X, µ), b ∈ FR(Γ, Y, ν) and b v a, then C(b) ≥ C(a). In particular, a ∼ =w b ⇒ C(a) = C(b), i.e., the cost is an invariant of weak isomorphism. (Note here that Γ is not assumed to be finitely generated.) To see this, consider the equivalence relations Ea , Eb induced by a, b, resp. Also fix a homomorphism π : X → Y of a to b. Given a graphing G of Eb (see Kechris-Miller [KM], Section 17) define a graph G 0 ⊆ Ea by xG 0 y ⇔ xEa y & π(x)Gπ(y). We claim that G 0 is a graphing of Ea . Indeed let γ a (x) = y, where γ ∈ Γ. Then γ b (π(x)) = π(y), so there is a G-path π(x) = z0 Gz1 . . . Gzn = π(y). As G ⊆ Eb , there are δ0 , . . . , δn−1 ∈ Γ such that δib (zi ) = zi+1 . Thus γ = δn−1 · · · δ0 . Then π(δ0a (x)) = δ0b (z0 ) = z1 , and so if x1 = δ0a (x), we have xG 0 x1 . Next π(δ1a (x1 )) = δ1b (z1 ) = z2 , so if x2 = δ1a (x1 ), x1 G 0 x2 . Continuing this way, we find a G 0 -path from x to y. Next note that if dG (y) denotes the degree of G at y, then dG 0 (x) ≤R dG (π(x)) (using the R freeness of the 1 0) = 1 0 d (x)dµ(x) ≤ dG (π(x))dµ(x) = action b). So C (E ) ≤ C (G µ a µ G 2 2 R 1 dG (y)dν(y) = Cµ (G). Since Cµ (Eb ) = inf G Cµ (G) (see Kechris-Miller 2 [KM], Section 18), it follows that C(b) ≥ C(a).
11. CHARACTERIZATIONS OF GROUPS WITH PROPERTY (T) AND HAP
79
Note that Q the fact mentioned in the proof of 10.12 is a special case of this, as b = n an ⇒ an v b, ∀n. When Γ is infinite and finitely generated, 10.14 shows that C is also an invariant of weak equivalence, but I do not know if this stronger fact holds for every infinite, countable Γ. 11. Characterizations of groups with property (T) and HAP (A) We first consider a characterization of groups with the Haagerup Approximation Property (HAP). We use Cherix et al. [CCJJV] as a standard reference for HAP. Moreover, other prerequisite material is developed in Appendices A–H. The groups Γ with HAP can be defined in several equivalent ways. We will use below the following ones: (i) Γ has HAP iff there is a unitary representation π of Γ such that 1Γ ≺ π and π is a c0 -representation. Recall that 1Γ is the trivial one-dimensional representation and 1Γ ≺ π is equivalent to the existence of non-0 almost invariant vectors, i.e., the existence of a sequence {vn } of unit vectors such that kπ(γ)(vn ) − vn k → 0, ∀γ ∈ Γ. Also π is a mixing representation or c0 -representation if the matrix coefficients hπ(γ)(v), vi → 0 as γ → ∞, for every v ∈ Hπ . (ii) Γ has HAP iff there is an orthogonal representation π of Γ (on a real Hilbert space) which has non-0 almost invariant vectors and is a c0 representation (as in (i)). It is clear that (ii) ⇒ (i) (using complexifications as in Appendix A). For the proof of (i) ⇒ (ii), let {vn } be as in i), and put ϕn (γ) = hπ(γ)(vn ), vn i. Then ϕn is positive-definite, ϕn (1) = 1, ϕn (γ) → 0 as γ → ∞, and moreover Re ϕn (γ) → 1 as n → ∞. Then ψn = Re ϕn is real positive-definite (see Berg-Christensen-Ressel [BCR], 3.1.18), ψn (1) = 1, ψn (γ) → 0 as γ → ∞ and ψn (γ) → 1 as n → ∞. Let (πn , Hn , wn ) be the orthogonal representation given by the GNS construction for ψn , so that wn is a cyclic vector and hπn (γ)(w L n ), wn i =Lψn (γ) (thus, in particular, wn is a unit vector). Let H = n Hn , π = n πn . Clearly hπ(γ)(wn ), wn i = ψn (γ) → 1 as n → ∞, thus kπ(γ)(wn ) − wn k → 0, and for v ∈ H, hπ(γ)(v), vi → 0 as γ → ∞, so (ii) holds. Theorem 11.1 (Jolissaint [CCJJV]). Let Γ be an infinite countable group. Then the following are equivalent: (i) Γ has the Haagerup property. (ii) Γ has a measure preserving, mixing action which is not E0 -ergodic. (iii) Γ has a free, measure preserving, mixing action which is not E0 ergodic. In particular, Γ does not have HAP iff MIX(Γ, X, µ) ⊆ E0 RG(Γ, X, µ). Proof. (ii) ⇒ (iii): If a ∈ A(Γ, X, µ) satisfies (ii) and b ∈ A(Γ, X, µ) is a free, mixing action of Γ (e.g., the shift of Γ on 2Γ ), then a × b satisfies (iii).
80
II. THE SPACE OF ACTIONS
(iii) ⇒ (i): Let a ∈ A(Γ, X, µ) satisfy (iii). Then if κa0 is the corresponding Koopman representation restricted to L20 (X, µ), κa0 is a c0 -representation and a is not E0 -ergodic, so 1Γ ≺ κa0 (see, e.g., 10.6 and the paragraph following it). (i) ⇒ (ii). Since Γ has HAP, there is an orthogonal representation π of Γ on H which has non-0 almost invariant vectors {vn } and is a c0 representation. By replacing π by infinitely many copies of it, we can assume that {vn } is orthonormal and of course H is infinite-dimensional. Consider the product space (X, ν) = (RN , µN ), with µ normalized Gaussian measure on R. Then there is an isomorphism θ : H → H :1: , where H :1: = hpn in∈N is the first chaos of L20 (X, ν, R) (with pn the projection function pn : RN → R). The representation π via θ gives rise to a measure preserving action a on (X, ν) such that if κa0 is the corresponding Koopman representation, then H :1: is κa0 -invariant and θ sends π to κa0 |H :1: . From the Wiener chaos decomposition, using the obvious fact that if π is a c0 -representation on H, then π n is a c0 -representation on H n , it follows that κa0 is a c0 -representation, so a is mixing. It remains to show that it is not E0 -ergodic. Consider the Polish space RepO (Γ, H :1: ) of orthogonal representations of Γ on H :1: , viewed as a closed subspace of O(H :1: )Γ , with O(H :1: ) the orthogonal group of H :1: (see Appendix A). Since O(H :1: ) can be identified with a closed subgroup of Aut(X, ν), RepO (Γ, H :1: ) is identified with a closed subspace of A(Γ, X, ν) by identifying any ρ ∈ RepO (Γ, H :1: ) with the unique action b ∈ A(Γ, X, ν) such that κb0 |H :1: = ρ. Moreover the conjugacy class of ρ in RepO (Γ, H :1: ) is contained, when we identify ρ and b, in the conjugacy class of b in Aut(Γ, X, µ). Now it is easy to check that the trivial representation π0 ∈ RepO (Γ, H :1: ) is contained in the closure of the conjugacy class of (the image under θ of) π in RepO (Γ, H :1: ) and, as the action corresponding to π0 is the trivial action iΓ ∈ A(Γ, X, ν), we have that iΓ ≺ a, therefore a 6∈ E0 RG(Γ, X, µ) (see 10.6). 2 Remark. More generally, the simple fact that is being used here is that if σ, τ ∈ RepO (Γ, H :1: ) and c, d ∈ A(Γ, X, ν) are the actions such that σ = κc0 |H :1: and τ = κd0 |H :1: , then σ ≺Z τ ⇒ c ≺ d, where σ ≺Z τ is the obvious orthogonal analog of the notion of weak containment in the sense of Zimmer (see the Remark after H.2), which is equivalent to the property that σ is in the closure of the conjugacy class of τ in RepO (Γ, H :1: ). There is also another way to see that a, in the proof of 11.1 (i) ⇒ (ii), is not E0 -ergodic (following the method of Connes-Weiss [CW]). What one needs to show is that if π is an orthogonal representation of Γ on H :1: , a is the action in A(Γ, X, ν) with π = κa0 |H :1: , and if π has non-0 almost invariant vectors {qn }, then a has non-trivial almost invariant sets.
11. CHARACTERIZATIONS OF GROUPS WITH PROPERTY (T) AND HAP
81
To prove this, put An = {x ∈ X : qn (x) ≥ 0}. Since every function in H :1: is a centered Gaussian random variable, clearly ν(An ) = 12 . We will verify that ν(γ a (An )∆An ) → 0, ∀γ ∈ Γ, or equivalently ν(γ a (An ) \ An ) → 0, ∀γ ∈ Γ. Fix γ ∈ Γ. Since κa0 (γ)(H :1: ) = H :1: , we have κa0 (γ)(qn ) = (cos an )qn + (sin an )rn , 0 ≤ an < 2π, where rn ∈ H :1: is a unit vector perpendicular to the unit vector qn . Then qn , rn are independent centered normalized Gaussian random variables. Clearly cos an = hπ(γ)(qn ), qn i → 1 as n → ∞. Now x ∈ γ · An \ An ⇔ [κa0 (γ)(qn )(x) ≥ 0 and qn (x) < 0] ⇔ [(cos an )qn (x) + (sin an )rn (x) ≥ 0 and qn (x) < 0], thus, since qn , rn are independent, normalized Gaussian variables, ν(γ · An \ An ) = µ2 ({(q, r) ∈ R2 : (cos an )q + (sin an )r ≥ 0 and q < 0}), so ν(γ · An \ An ) → 0. (B) We finally prove the following characterization of groups with property (T), which preceded historically the characterization 11.1 of groups with HAP. Recall that Γ has property (T) if for every unitary representation π of Γ, 1Γ ≺ π implies 1Γ ≤ π. Theorem 11.2 (Connes-Weiss [CW], Schmidt [Sc3], Glasner-Weiss [GW]). Let Γ be an infinite countable group. Then the following are equivalent: (i) Γ does not have property (T). (ii) There is a measure preserving, ergodic action of Γ which is not E0 ergodic. (iii) There is a free, measure preserving, ergodic action of Γ which is not E0 -ergodic. (iv) There is a measure preserving, weak mixing action of Γ which is not E0 -ergodic. (v) There is a free, measure preserving, weak mixing action of Γ which is not E0 -ergodic. In particular, Γ has property (T) iff WMIX(Γ, X, µ) ⊆ E0 RG(Γ, X, µ) iff ERG(Γ, X, µ) = E0 RG(Γ, X, µ). Proof. (v) ⇒ (iv) ⇒ (ii), (v) ⇒ (iii) ⇒ (ii) are obvious. (ii) ⇒ (i). Suppose a is a measure preserving, ergodic action of Γ on (X, µ) which is not E0 -ergodic. Then if κa0 is the corresponding Koopman representation on L20 (X, µ), 1Γ ≺ π0a (since the action is not E0 -ergodic) and 1Γ 6≤ κa0 , since the action is ergodic. So Γ does not have property (T). (i) ⇒ (iv) ⇒ (v). As in the proof of 11.1, it is enough, assuming Γ does not have property (T), to find an action which is measure preserving, weak mixing and not E0 -ergodic. By the definition of property (T), there is a unitary representation ρ of Γ on Hρ such that 1Γ ≺ ρ but 1Γ 6≤ ρ. Let Γ = {γ1 , γ2 , . . . } and, since 1Γ ≺ ρ,
82
II. THE SPACE OF ACTIONS
fix a sequence of unit vectors v1 , v2 , . . . such that kρ(γ)(vn ) − vn k2 < γ ∈ {γ1 , . . . , γn }. Let X ϕ(γ) = (nkρ(γ)(vn ) − vn k2 ) < ∞.
1 2n
for
n
Since kρ(γ)(vn )−vn k2 = 2−2 Re hρ(γ)(vn ), vn i, clearly ϕn (γ) = kρ(γ)(vn )− vn k2 is negative-definite and so is ϕ (also note that ϕ(1) = 0). Recall that a real functionPθ on Γ is negative-definite if θ(γ) = θ(γ −1 ) and moreover we have that 1≤i,j≤m αi αj θ(γi−1 γj ) ≤ 0, for all γ1 , . . . , γm ∈ Γ and P α1 , . . . , αm ∈ R with m i=1 αi = 0. We next show that ϕ is unbounded. First notice that for each n, there is δn such that kρ(δn )(vn ) − vn k2 > 41 . Indeed, otherwise kρ(γ)(vn ) − vn k ≤ 1 2 , ∀γ ∈ Γ, so the closure of the convex hall of {ρ(γ)(vn ) : γ ∈ Γ} is contained in the ball of radius 21 around vn . Then if v¯n is the unique element of least norm in this closed convex set, clearly v¯n 6= 0 and v¯n is Γ-invariant, contradicting that 1Γ 6≤ ρ. It follows that ϕ(δn ) > n4 , so ϕ is unbounded. Consider now the real positive-definite function ψn = e−ϕ/n , and let (πn , Hn , wn ) be the orthogonal representation of Γ given by the GNS construction for ψn , so that wn is a cyclic vector and hπn (γ)(wn ), wL n i = ψn (γ). (In particular, as ψ (1) = 1, w is a unit vector.) Let H = n n n Hn , π = L n πn . As ψn (γ) → 0, for n → ∞, the sequence {wn } witnesses that H has non-0 almost invariant vectors. Next as in the proof of (i) ⇒ (ii) for 11.1, whose notation we use below, let a ∈ A(Γ, X, ν) be the corresponding action. As in that proof, a is not E0 -ergodic, so it remains to show that it is weak mixing, i.e., that the corresponding Koopman representation κa0 is weak mixing. We first recall the following equivalent characterizations for a unitary representation σ of Γ on Hσ to be weak mixing (see Bergelson-Rosenblatt [BR], 1.9 and 4.1, Glasner [Gl2], 3.2, 3.5 and 1.52): (a) σ has no non-0 finite-dimensional subrepresentations, (b) ∀ξ1 , . . . , ξn ∈ Hσ ∀ > 0∃γ ∈ Γ∀i, j(|hσ(γ)(ξi ), ξj i| < ), (c) If M is the mean on the weakly almost periodic functions on Γ (see Glasner [Gl2], 1.51), then for any v ∈ Hσ , if ψ(γ) = hσ(γ)(v), vi, then M(|ψ|2 ) = 0. Recall here that every positive-definite function is weakly almost periodic and ψ, |ψ|2 as above are positive-definite. Moreover, for any positive-definite function ω, m X X M(ω) = inf{ αi αj ω(δi−1 δj ) : αi ≥ 0, αi = 1, δ1 , . . . , δm ∈ Γ} i=1
(see Godement [Go], THEOREME 14). Using (b) for σ = κa0 and the Wiener chaos decomposition, it is easy to see that it is enough to show that κa0 |HC:1: is weak mixing, which (by either (a) or (c)) is reduced to showing that each πn + i · πn is weak mixing. For this it is clearly enough to show that M(ψn2 ) = M(e−2ϕ/n ) = 0. This is
11. CHARACTERIZATIONS OF GROUPS WITH PROPERTY (T) AND HAP
83
a general fact about unbounded negative-definite functions (see Jolissaint [Jo1], 4.4 or [Jo3], 2.1) but it is easy to adapt the proof in our case without using any general theory. We will show that M(e−ϕ ) = 0, the proof for e−2ϕ/n being identical. First note that we can represent ϕ in the form ϕ(γ −1 δ) = kf (γ) − f (δ)k2 for someL function f : Γ → F, F a Hilbert space, with f (1) = 0. Indeed, let F =L n Hρ be the direct sum of infinitely many copies of Hρ and put √ f (γ) = n n(ρ(γ)(vn ) − vn ). In particular, ϕ(γ) = kf (γ)k2 , therefore ϕ(γ −1 δ) = kf (γ) − f (δ)k2 ≥ (kf (γ)k − kf (δ)k)2 p p = | ϕ(γ) − ϕ(δ)|2 . Recalling that m X X −1 M(e−ϕ ) = inf{ αi αj e−ϕ(δi δj ) : αi ≥ 0, αi = 1, δ1 , · · · , δm ∈ Γ}, i=1
P 1 1 2 fix > 0 and find m large enough so that if ai = m , then m i=1 αi =p m < /2. Then inductively define δ1 = 1, . . . , δm , so that, if 1 ≤ j < i ≤ m, ϕ(δi ) > q p −1 log( 2 ) + ϕ(δj ), and thus ϕ(δi−1 δj ) > log(2/), e−ϕ(δi δj ) < /2, and P −1 therefore αi αj e−ϕ(δi δj ) < . 2 Remark. In the proof of (i) ⇒ (iv) in 11.2, we have started with a representation ρ of a countable group Γ such that 1Γ ≺ ρ and 1Γ 6≤ ρ (i.e., ρ is ergodic). We then constructed a real negative-definite ϕ which is unbounded and then considered the real positive-definite ψn = e−ϕ/n and the corresponding orthogonal representation πn and finally the direct sum L π = n πn . If τ = π ⊕ i · π is the complexification of π, then 1Γ ≺ τ and the unboundedness of ϕ was used to show that πn ⊕ i · πn is weak mixing, thus so is τ . So by starting with an ergodic ρ with 1Γ ≺ ρ one ends up with a weak mixing τ with 1Γ ≺ τ . (C) Schmidt [Sc4] calls groups Γ for which WMIX(Γ, X, µ) = ERG(Γ, X, µ) groups with property (W). He characterizes infinite such groups as those that admit no non-trivial finite-dimensional unitary representations (such groups are called minimally almost periodic). Dye [Dy] has shown that there are amenable groups with property (W). Schmidt [Sc4] also considers groups with property (M) and property (S), characterized by the equalities MMIX(Γ, X, µ) = ERG(Γ, X, µ), and MIX(Γ, X, µ) = ERG(Γ, X, µ), resp. It is apparently unknown whether (countable infinite) such groups exist.
84
II. THE SPACE OF ACTIONS
12. The structure of the set of ergodic actions (A) First we note the following fact. Proposition 12.1. Let Γ be a countable group. The sets ERG(Γ, X, µ), WMIX(Γ, X, µ) are Gδ in the weak topology of A(Γ, X, µ). Proof. First consider ERG(Γ, X, µ). One way to prove this is to use Section 10, (B). Take (X, µ) = (T, λ). The ergodic (relative to the shift s) measures in SIM(Γ) are the extreme points of SIM(Γ), thus the set ESIM(Γ) of ergodic shift-invariant measures is Gδ and so is ESIMλ (Γ) = SIMλ (Γ)∩ ESIM(Γ). Since Φλ is continuous and Φ−1 λ (ESIMλ (Γ)) = ERG(Γ, T, λ) the proof is complete. An alternative way to prove this is as follows: The action a ∈ A(Γ, X, µ) is ergodic iff the corresponding Koopman representation κa0 on L20 (X, µ) is ergodic, i.e., for every v ∈ L20 (X, µ), if ϕv (γ) = hκa0 (γ)(v), vi, then M(ϕv ) = 0, where M is the mean on the weakly almost periodic functions on Γ (see Glasner [Gl2], 3.2, 3.9, 3.10). Now recall that M(ϕv ) = inf{
X
αi αj ϕv (γi−1 γj ) : αi ≥ 0,
n X
αi = 1, γ1 , . . . , γn ∈ Γ}.
i=1
Moreover, it is clear that we can restrict v here to any countable dense set D ⊆ L20 (X, µ). Thus, restricting below to rationals, n , (γ )n a ∈ ERG(Γ, X, µ) ⇔ ∀v ∈ D∀ i i=1 i=1 P> 0∃(αi )P [αi ≥ 0, αi = 1, αi αj ϕv (γi−1 γj ) < ],
so clearly ERG(Γ, X, µ) is Gδ . A similar argument works for WMIX. By Bergelson-Rosenblatt [BR], 4.1, we have a ∈ WMIX(Γ, X, µ) ⇔ ∀v1 , . . . , vn ∈ D∀∃γ(|hκa0 (γ)(vi ), vi i| < , ∀i ≤ n). Alternatively we can use the fact that a ∈ WMIX(Γ, X, µ) ⇔ a × a ∈ ERG(Γ, X × X, µ × µ) and a 7→ a × a is continuous in the weak topology.
2
Note also that ERG(Γ, X, µ) has no interior in w. To see this take a non-ergodic action b and any a ∈ A(Γ, X, µ). We have seen in 10.4 that a is in the closure of the set of isomorphic copies of a × b, which is clearly non-ergodic. Thus ERG(Γ, X, µ) has no interior. (B) We have the following important dichotomy, originally formulated in Glasner-Weiss [GW] in terms of the structure of the set of ergodic measures in SIM(Γ). Theorem 12.2 (Glasner-Weiss [GW]). i) If a countable group Γ has property (T), then the set of ergodic actions ERG(Γ, X, µ) is closed in the weak topology of A(Γ, X, µ) (and has no interior).
12. THE STRUCTURE OF THE SET OF ERGODIC ACTIONS
85
ii) If a countable group Γ does not have property (T), then ERG(Γ, X, µ) is dense Gδ in the weak topology of A(Γ, X, µ). Proof. i) Given a ∈ A(Γ, X, µ) consider the corresponding Koopman unitary representation κa0 on L20 (X, µ). Then a ∈ ERG(Γ, X, µ) ⇔ κa0 has no invariant non-0 vectors. Recall that a Kazhdan pair for Γ is a pair (Q, ), where Q ⊆ Γ, Q finite, > 0, such that in any unitary representation π of Γ, if there is a vector v with kπ(γ)(v) − vk < kvk, ∀γ ∈ Q, then there is a non-0 invariant vector. It is well-known (see Bekka-de la Harpe-Valette [BdlHV]) that Γ has property (T) iff it admits a Kazhdan pair. Then for a Kazhdan pair (Q, ), a 6∈ ERG(Γ, X, µ) ⇔ ∃v ∈ L20 (X, µ)∀γ ∈ Q kκa0 (γ)(v) − vk < kvk, which clearly shows that A(Γ, X, µ) \ ERG(Γ, X, µ) is open in w. ii) Using 10.9, it is enough to show that ERG(Γ, X, µ) is dense in the set FED(Γ, X, µ). Again for that it is enough to show that if a1 , a2 ∈ ERG(Γ, X, µ), λ1 +λ2 = 1, 0 ≤ λi ≤ 1, then λ1 a1 +λ2 a2 ≺ b for some ergodic b. Because then by induction on n it is easy to see (using Pn the paragraph after 10.9) that if ai ∈ ERG(Γ, X, µ), 0 ≤ λi ≤ 1, then i=1 λi ai ≺ b for some ergodic b. First notice that there is an ergodic a such that a1 , a2 ≺ a. This follows from 10.4, by taking ρ to be an ergodic joining of a1 , a2 (see also the proof of 13.1). So λ1 a1 + λ2 a2 ≺ λ1 a + λ2 a, thus it is enough to show for ergodic a and 0 ≤ λ1 , λ2 ≤ 1, λ1 + λ2 = 1, that λ1 a + λ2 a ≺ b for some ergodic b. Since Γ does not have property (T), by 11.2 there is a0 ∈ WMIX(Γ, X, µ) \ E0 RG(Γ, X, µ). We will in fact show that λ1 a + λ2 a ≺ a0 × a. Since a0 is not E0 -ergodic, find a sequence A0n of Borel sets such that µ(A0n ) = λ1 (so µ(X \ A0n ) = λ2 ) and µ(γ a0 (A0n )∆A0n ) → 0, ∀γ ∈ Γ (i.e., {A0n } is almost invariant for a0 ; see Jones-Schmidt [JS] or Hjorth-Kechris [HK3], A2.2, A2.3). Let Y = X1 t X2 be the disjoint union of two copies of X and let ν = λ1 µ1 + λ2 µ2 , where µi is the copy of µ on Xi . Thus ν(Xi ) = λi . Let c be the union of the copy of a on X1 and the copy of a on X2 . Thus c ∼ = λ1 a + λ2 a. We will check that c ≺ a0 × a. Indeed given Borel sets (j) A1 , . . . , Am in Y , let Ai = Ai ∩ Xj ⊆ X. Put (n)
Bi
(1)
(2)
= (A0n × Ai ) ∪ ((X \ A0n ) × Ai ).
Fix F ⊆ Γ finite and > 0. Then we claim that if n is large enough (n)
(n)
|ν(γ c (Ai ) ∩ Ak ) − (µ × µ)(γ a0 ×a (Bi ) ∩ Bk )| < . Indeed
(1)
(1)
ν(γ c (Ai ) ∩ Ak ) = λ1 µ(γ a (Ai ) ∩ Ak )+ (2) (2) λ2 µ(γ a (Ai ) ∩ Ak )
86
II. THE SPACE OF ACTIONS
and if n is large enough (n)
γ a0 ×a (Bi ) (1)
(2)
is very close to (A0n ×γ a (Ai ))∪((X \A0n )×γ a (Ai )) in the measure algebra (n) (n) (1) (1) of µ × µ, thus γ a0 ×a (Bi ) ∩ Bk is very close to (A0n × (γ a (Ai ) ∩ Ak )) ∪ (2) (2) ((X \ A0n ) × (γ a (Ai ) ∩ Ak )), which has µ × µ-measure equal to (1)
(1)
(2)
(2)
λ1 µ(γ a (Ai ) ∩ Ak ) + λ2 µ(γ a (Ai ) ∩ Ak ), so we are done.
2
Remark. A similar argument as in the proof of 12.2, i) above can be also given to show the result of Glasner-Weiss that if Γ has property (T), then the set ESIM(Γ) of ergodic measures in SIM(Γ) (see Section 10, (B)) is closed in SIM(Γ). One only needs to notice that in this proof v can be restricted to any dense set in L20 (X, µ) and that if X is compact, the continuous functions in L20 (X, µ) are dense in L20 (X, µ). Remark. Note that 12.2, ii) is equivalent to the fact that ESIM(Γ) is dense Gδ in SIM(Γ), which is what is proved in Glasner-Weiss [GW]. Indeed, using the Glasner-King [GK] correspondence described in Section 10, (B), we have that (h, a) ∈ Φ−1 (ESIM(Γ)) ⇔ a ∈ ERG(Γ, T, λ) and so ESIM(Γ) is dense Gδ in SIM(Γ) iff Φ−1 (ESIM(Γ)) is dense Gδ in H × A(Γ, T, λ) iff ERG(Γ, T, λ) is dense Gδ in A(Γ, T, λ). (C) On the other hand, for any group Γ, there are very few E0 -ergodic actions. Proposition 12.3. For any countable group Γ, the set E0 RG(Γ, X, µ) is meager in (A(Γ, X, µ), w). Proof. The set of actions I(Γ, X, µ) which admit almost invariant sets of measure 1/2, i.e., ∀F finite ⊆ Γ∀ > 0∃A(µ(A) = 21 & ∀γ ∈ F µ(γ · A∆A) < ) is a Gδ set. Let {an }∞ X, µ) and let a0 ∈ A(Γ, X, µ) n=1 be dense in A(Γ,Q have an invariant set of measure 12 . If a∞ = ∞ n=0 an , then the isomorphic copies of a∞ are dense in (A(Γ, X, µ), w), by 10.4, and clearly they belong to I(Γ, X, µ), so I(Γ, X, µ) is dense Gδ . Clearly I(Γ, X, µ) ∩ E0 RG(Γ, X, µ) = ∅ (see the paragraph preceding 10.6), so E0 RG(Γ, X, µ) is meager. 2 Thus we also have another dichotomy. Corollary 12.4. i) If a countable group Γ has property (T), the set of ergodic but not E0 -ergodic actions is empty. ii) If a countable group Γ does not have property (T), the set of ergodic but not E0 -ergodic actions is dense Gδ in (A(Γ, X, µ), w). Proof. i) follows from 11.2. For ii) notice that ERG(Γ, X, µ) \ E0 RG(Γ, X, µ) = ERG(Γ, X, µ) ∩ I(Γ, X, µ) (as in the proof of 12.3) and use 12.2. We finally note the following corollary of 12.2.
2
12. THE STRUCTURE OF THE SET OF ERGODIC ACTIONS
87
Corollary 12.5. Let Γ be a countable group. Then the following are equivalent: (i) Γ has property (T), (ii) ERG(Γ, X, µ) is closed in (A(Γ, X, µ), w), (iii) ERG(Γ, X, µ) is Fσ in (A(Γ, X, µ), w). Proof. (i) ⇒ (ii) is 12.2, i) and (ii) ⇒ (iii) is obvious. Assume now (iii). Then as the complement of ERG(Γ, X, µ) is dense Gδ in A(Γ, X, µ), 12.2 ii) shows that Γ has property (T), i.e., (i) holds. 2 This corollary provides an abstract “descriptive” way of thinking about property (T) for a countable group Γ, in terms of the structure of the set of its ergodic actions. It has the following immediate consequence: Suppose UΓ ⊆ A(Γ, X, µ) is a Gδ property (set) of measure preserving actions of Γ (in the weak topology), which is satisfied by all non-ergodic actions, i.e., A(Γ, X, µ) \ ERG(Γ, X, µ) ⊆ UΓ . Then Γ has property (T) if UΓ implies (thus is equivalent to) non-ergodicity. For example, fix finite Q ⊆ Γ, > 0, and call a set A ∈ MALGµ , (Q, )-invariant for an action a, if ∀γ ∈ Q(µ(γ a (A)∆A) < µ(A)(1 − µ(A)). Recall that (Q, ) is a Kazhdan pair for Γ if for every unitary representation π : Γ → U (H), if there is v ∈ H which is (Q, )-invariant, i.e., kπ(γ)(v) − vk < kvk, ∀γ ∈ Q, then there exists v 6= 0 which is invariant, i.e, π(γ)(v) = v, ∀γ ∈ Γ. If (Q, ) is a Kazhdan pair for Γ and a ∈ A(Γ, X, µ) admits a non-trivial (Q, 2 )-invariant set A, then in the Koopman representation κa0 we have, for v = χA − µ(A), kκa0 (γ)(v) − vk2 = µ(γ a (A)∆A), kvk2 = µ(A)(1 − µ(A)), thus kκa0 (γ)(v) − vk < kvk, ∀γ ∈ Q, and so κa0 admits a non-0 invariant vector, therefore a is not ergodic. So this leads to the following conclusion. Corollary 12.6. Let Γ be a countable group. Then the following are equivalent: (i) Γ has property (T), (ii) There is finite Q ⊆ Γ, > 0, such that every a ∈ A(Γ, X, µ) which admits a (Q, )-invariant set has a non-trivial invariant set, i.e., is not ergodic. Proof. Let UΓ,Q, = {a ∈ A(Γ, X, µ) : a admits a (Q, )-invariant set}. Then UΓ,Q, is open in (A(Γ, X, µ), w) and A(Γ, X, µ) \ ERG(Γ, X, µ) ⊆ UΓ,Q, . So if (ii) holds for some (Q, ), Γ has property (T). Conversely, if Γ has property (T) with Kazhdan pair (Q, ), by the paragraph preceding 12.6, (Q, 2 ) satisfies (ii). 2
88
II. THE SPACE OF ACTIONS
(D) In Glasner [Gl2], 13.20 a characterization of groups failing to have property HAP was obtained, analogous to that for property (T) groups that was established in Glasner-Weiss [GW]. This characterization asserted that the closure of the mixing shift-invariant measures on TΓ is contained in the ergodic shift-invariant measures. By a combination of the arguments used in the proof of 12.2 and 10.4, one can obtain a similar characterization for the space of actions. Theorem 12.7. Let Γ be an infinite countable group. Then Γ does not have HAP iff MIX(Γ, X, µ) ⊆ ERG(Γ, X, µ) (where closure is in the weak topology). Proof. ⇒: If Γ does not have HAP, then MIX(Γ, X, µ) ⊆ {a ∈ A(Γ, X, µ) : 1Γ 6≺ κa0 }. Assume now, towards a contradiction, that an ∈ MIX(Γ, X, µ) and an → Q a 6∈ ERG(Γ, X, µ). Let b = n an . Then, by 10.4, an ≺ b, ∀n, so a ≺ b and thus κa0 ≺ κb0 . Now, as a is not ergodic, 1Γ ≤ κa0 , so 1Γ ≺ κb0 , contradicting the fact that b is mixing (as a product of mixing actions.) ⇐: Now assume that Γ has HAP. Then, by 11.1, there is an action a0 ∈ MIX(Γ, X, µ) \ E0 RG(Γ, X, µ) and thus, for the action a0 , there is an almost invariant sequence of Borel sets {An } in X with µ(An ) = 12 . Then, as in the proof of 12.2, ii), we see that b = 12 a0 + 21 a0 ≺ a0 × a0 . Clearly b is not ergodic, a0 × a0 is mixing, and so b ∈ MIX(Γ, X, µ), thus MIX(Γ, X, µ) 6⊆ ERG(Γ, X, µ). 2 Note also that a similar argument shows that if Γ has HAP, then the space MIX(Γ, X, µ) is closed under convex combinations. Indeed let a1 , a2 ∈ MIX(Γ, X, µ), 0 ≤ λi ≤ 1, λ1 + λ2 = 1, (n)
(n)
and let ai ∈ MIX(Γ, X, µ), ai → ai (in the weak topology). Then if (n) (n) (n) (n) a(n) = a1 × a2 , λ1 a1 + λ2 a2 ≺ λ1 a(n) + λ2 a(n) ≺ a0 × a(n) , for an appropriate a0 ∈ MIX(Γ, X, µ) \ E0 RG(Γ, X, µ), and, as a0 , a(n) , a0 × a(n) ∈ MIX(Γ, X, µ), we have λ1 a1 + λ2 a2 ∈ MIX(Γ, X, µ). Thus an infinite Γ has HAP iff MIX(Γ, X, µ) is closed under convex combinations. (For the analogous result concerning shift-invariant measures on TΓ , see Glasner [Gl2], 13.21.) We also have the following characterization of property (T) groups. Theorem 12.8. Let Γ be an infinite countable group. Then Γ has property (T) iff WMIX(Γ, X, µ) ⊆ ERG(Γ, X, µ) (where closure is in the weak topology) iff WMIX(Γ, X, µ) is closed in the weak topology. Proof. The first equivalence is proved as in 12.7, using also 12.2. If Γ has property (T), then ERG(Γ, X, µ) is closed. But a ∈ WMIX(Γ, X, µ) ⇔ a × a ∈ ERG(Γ, X × X, µ × µ)
12. THE STRUCTURE OF THE SET OF ERGODIC ACTIONS
89
and a 7→ a × a is continuous in the weak topology, so WMIX(Γ, X, µ) is also closed. Clearly if WMIX(Γ, X, µ) is weakly closed, WMIX(Γ, X, µ) = WMIX(Γ, X, µ) ⊆ ERG(Γ, X, µ). 2 Again we also have that Γ does not have property (T) iff WMIX(Γ, X, µ) is closed under convex combinations. In the recent paper Kerr-Pichot [KP] the authors independently note that if Γ has property (T), then WMIX(Γ, X) is closed in the weak topology. Moreover they establish the following strengthening of 12.2, (ii), which answers a question of Bergelson-Rosenblatt [BR], p. 80. Theorem 12.9 (Kerr-Pichot [KP]). If a countable group Γ does not have property (T), then WMIX(Γ, X, µ) is dense Gδ in the weak topology of A(Γ, X, µ). Proof (sketch). Fix a sequence of Borel finite partitions P1 , P2 , . . . of X, each refining the preceding one, such that the union of the finite Boolean algebras generated by each Pn is dense in the measure algebra. Then the set of weak mixing actions is equal to the intersection of the sets 1 )}. m So, by the Baire Category Theorem, it is enough to show that for each finite Borel partition P of X and > 0, the open set Wn,m = {a ∈ A(Γ, X, µ) : ∃γ∀A, B ∈ Pn (|µ(γ a (A) ∩ B) − µ(A)µ(B)| <
{a ∈ A(Γ, X, µ) : ∃γ∀A, B ∈ P(|µ(γ a (A) ∩ B) − µ(A)µ(B)| < )} is dense in the weak topology. By considering refinements of partitions, it is enough to show that given a ∈ A(Γ, X, µ), a finite Borel partition Q of X, 1 ∈ F ⊆ Γ finite and δ > 0, there is b ∈ A(Γ, X, µ) such that µ(γ b (A)∆γ a (A)) < δ, ∀A ∈ Q, ∀γ ∈ F and ∃γ(|µ(γ b (A) ∩ B) − µ(A)µ(B)| < δ, ∀A, B ∈ Q). Since Γ does not have property (T), 11.2 shows that WMIX(Γ, X, µ) \ E0 RG(Γ, X, µ) 6= ∅ and so we can find c ∈ WMIX(Γ, X, µ) such that for each > 0, finite F ⊆ Γ, there is a set A with µ(A) = 1/2 and µ(γ c (A)∆A) < , ∀γ ∈ F (see, e.g., Hjorth-Kechris [HK3], A2.2). By considering the weak Q mixing product cn and the partition { 1≤i≤n Ap(i) : p ∈ {−1, 1}n }, where A1 = A, A−1 = X \ A, it follows that for each > 0, n ≥ 1, F ⊆ Γ finite, we can find d ∈ WMIX(Γ, X, µ) and a Borel partition R of X into 2n pieces of equal measure such that µ(γ d (A)∆A) < , ∀γ ∈ F, ∀A ∈ R. (Alternatively, one can use the argument in the second part of the proof of 10.6 to show that we can actually take d = c.) Fix now a ∈ A(Γ, X, µ), 1 ∈ F ⊆ Γ finite, a finite Borel partition Q of X and δ > 0, in order to find b ∈ A(Γ, X, µ) such that µ(γ b (A)∆γ a (A)) < δ, ∀A ∈ Q, ∀γ ∈ F and ∃γ[|µ(γ a (A) ∩ B) − µ(A)µ(B)| < δ, ∀A, B ∈ Q]. First let n, be such that 2−n < δ, 2n < δ. Then find d, R for these , n, F . Say R = {R1 , . . . , Rm }, with m = 2n . Consider the product space
90
II. THE SPACE OF ACTIONS
(X m+1 , µm+1 ) and for B ⊆ X, let ˆ= B
m [
(Ri × X × · · · × B × · · · × X),
i=1
where B appears in the (i + 1)th coordinate in Ri × X × · · · × B × · · · × X. ˆ = µ(B), and B 7→ B ˆ preserves intersections and unions. Clearly µm+1 (B) m If a ˆ is the action iΓ × a , where iΓ is the trivial action of Γ on X, then a \ ˆ for any Borel set B ⊆ X and γ ∈ Γ. It follows that there γ (B) = γ aˆ (B), ˆ = γ a (A), ∀γ ∈ is a isomorphism ϕ : (X m+1 , µm+1 ) → (X, µ) with ϕ(γ aˆ (A)) m F, A ∈ Q. Let now e = d × a and b be the isomorphic copy of e on (X, µ) induced by ϕ. Then this b works. 2 Corollary 12.10. A countable group Γ does not have property (T) iff ERG(Γ, X, µ) is dense in the weak topology of A(Γ, X, µ) iff WMIX(Γ, X, µ) is dense in the weak topology of A(Γ, X, µ). From 12.7 it follows that for every Γ that does not have HAP, the set MIX(Γ, X, µ) is not dense in A(Γ, X, µ). In an earlier version of this work it was asked whether for any infinite group with HAP, MIX(Γ, X, µ) is dense in A(Γ, X, µ). This has now been answered by Hjorth. Theorem 12.11 (Hjorth [Hj4]). An infinite countable group Γ has the HAP iff MIX(Γ, X, µ) is dense in the weak topology of A(Γ, X, µ). I do not know if there are Γ for which MIX(Γ, X, µ) is comeager in the weak topology of A(Γ, X, µ). 13. Turbulence of conjugacy in the ergodic actions (A) We will now consider the conjugacy action of Aut(X, µ) on the space ERG(Γ, X, µ). First we show that this action has a dense orbit. Theorem 13.1. For any infinite countable group Γ, there is a free action in the space ERG(Γ, X, µ) with dense conjugacy class in (ERG(Γ, X, µ), w). In particular the set of free actions in ERG(Γ, X, µ) is dense Gδ in the space (ERG(Γ, X, µ), w). Proof. The proof is a modification of that of 10.7. Fix a dense set {an } in (ERG(Γ, X, µ), w), including a free action. For each n also fix a set Xn ⊆ X with µ(Xn ) = 1, such that Xn is an -invariant and µ is the Q unique ergodic, invariant measure for the action an restricted to Xn . Let a = n an . By the Q ergodic decomposition there is an ergodic, invariant measure ρ for a with ρ( n Xn ) = 1 and a is free ρ-a.e. Then if πn : X N → X denotes the nth projection, πn ({xiQ }) = xn , clearly Q (πn )∗ ρ = µ. So ρ is a joining of {an }. Then, by 10.4, an ≺ ( n an )ρ and ( n an )ρ is free. 2 It is clear that for any group Γ, if a ∈ ERG(Γ, X, µ) has dense conjugacy class in the weak topology of the space of ergodic actions, then a is free (since if a is not free, then for some γ ∈ Γ \ {1}, < 1, δu (γ a , 1) ≤ and
13. TURBULENCE OF CONJUGACY IN THE ERGODIC ACTIONS
91
closed balls in δu are weakly closed). Foreman-Weiss [FW] show that if Γ is amenable, then the conjugacy class of every free ergodic action is dense in (FRERG(Γ, X, µ), u) and thus it is dense in (ERG(Γ, X, µ), w). However, as we will now see in 13.2 below, this first property can only hold for amenable Γ. I do not know a characterization of the class of free ergodic actions that have weakly dense conjugacy class in ERG(Γ, X, µ) (for non-amenable Γ). Proposition 13.2. Let Γ be an infinite countable group. Let sΓ be the shift action of Γ on (X, µ), where X = 2Γ and µ is the usual product measure. Then the following are equivalent: (i) Γ is amenable. (ii) Every action in FRERG(Γ, X, µ) has dense conjugacy class in the space (A(Γ, X, µ), w). (iii) sΓ has dense conjugacy class in (A(Γ, X, µ), w), i.e., a ≺ sΓ for every a ∈ A(Γ, X, µ). (iv) iΓ ≺ sΓ . (v) There is an action in FRERG(Γ, X, µ) with dense conjugacy class in the space (FRERG(Γ, X, µ), u). (vi) Every action in FRERG(Γ, X, µ) has dense conjugacy class in the space (FRERG(Γ, X, µ), u). Proof. Let s = sΓ . (i) ⇒ (vi) is the aforementioned result of ForemanWeiss [FW], and clearly (vi) ⇒ (v). Next we claim that (v) ⇒ (vi). To see this, we note that the conjugacy action of Aut(Γ, X, µ) on A(Γ, X, µ) is an action by isometries for the metric δΓ,u . From this it is easy to see that the uniform closures of the conjugacy classes in the space A(Γ, X, µ) form a partition of this space (see, e.g., the proof of 23.1 below). So if some conjugacy class is dense in (FRERG(Γ, X, µ), u), every conjugacy class is dense in this space. We now show that (vi) ⇒ (ii). Indeed, if (vi) holds, then Γ cannot have property (T) by 14.2 and the fact that not all free ergodic actions of Γ are conjugate, which follows, for instance, from 13.7. But then by 13.1 and 12.2, ii), we obtain immediately (ii). Clearly (ii) ⇒ (iii) ⇒ (iv). Finally to see that (iv) ⇒ (i) note that iΓ κ0 = ∞ · 1Γ and κs0 ≺ λΓ (see Bekka-de la Harpe-Valette [BdlHV], E.3.5), so if iΓ ≺ sΓ , we have 1Γ ≺ λΓ and thus Γ is amenable (see, e.g., Zimmer [Zi1] or Bekka-de la-Harpe-Valette [BdlHV], G.3.2). 2 Remark. I would like to thank Eli Glasner for pointing out an inaccuracy in an earlier version of this proposition. Remark. In Ageev [A3] it is shown that if an infinite Γ is amenable, then every action in FR(Γ, X, µ) has dense conjugacy class in the space (A(Γ, X, µ), w), thus, using also 13.2, Γ is amenable iff every action in FR(Γ, X, µ) has dense conjugacy class in (A(Γ, X, µ), w). This answers a question in the above paper [A3].
92
II. THE SPACE OF ACTIONS
Remark. Note that by 12.2 and 13.1, Γ does not have property (T) iff there is an action in FRERG(Γ, X, µ) with dense conjugacy class in the space (A(Γ, X, µ), w). Remark. One can also ask if the weakening of (iii), where we require only that a ≺ sΓ for every a ∈ ERG(Γ, X, µ), is enough to imply that Γ is amenable. I do not know the answer but one can see that this weakening implies that Γ does not contain F2 . To prove this, we consider cases: If Γ does not have property (T), then, by 12.2, ii), a ≺ sΓ for all a ∈ A(Γ, X, µ), so Γ is amenable. So we can assume that Γ has property (T). We will then show that every infinite index subgroup ∆ ≤ Γ must be amenable. This implies that F2 ≤ Γ fails, since otherwise F2 would have finite index and thus property (T) (see Bekka-de la Harpe-Valette [BdlHV], 1.7.1), which is of course a contradiction. So fix ∆ ≤ Γ of infinite index, let X = Γ/∆ and consider the translation action of Γ on X : γ · δ∆ = γδ∆. Then let a be the shift action of Γ on 2X : a(γ, f )(x) = f (γ −1 · x), for γ ∈ Γ, f ∈ 2X , x ∈ X. One can identify 2X with the compact Cantor group ZX 2 , so this is an action by automorphisms on this group. Using 4.3 of Kechris [Kec4], we can see that (since X is infinite) a is ergodic, so a ≺ sΓ and therefore κa0 ≺ κs0Γ ≺ λΓ , i.e., the action a is tempered, in the terminology of [Kec4]. Using then 4.6 of [Kec4], we see that the stabilizer of any point in the action of Γ on X is amenable, i.e., ∆ is amenable. To summarize, we have seen that if Γ is such that it satisfies ∀a ∈ ERG(Γ, X, µ)(a ≺ sΓ ) but Γ is not amenable, then Γ must have property (T) and every infinite index subgroup of Γ must be amenable, so that F2 6≤ Γ. We also note that it must have the following property: Γ is minimally almost periodic, i.e., has no non-trivial finite-dimensional unitary representations. Indeed, this property is equivalent to the property that WMIX(Γ, X, µ) = ERG(Γ, X, µ) (see Schmidt [Sc4] and Section 11, (C)). If this failed for a Γ as above, then there is a ∈ ERG(Γ, X, µ) \ WMIX(Γ, X, µ), so a ≺ sΓ and thus κa0 ≺ κs0Γ ≺ λΓ and κa0 is not weak mixing, so it has a non-0 finitedimensional subrepresentation π. Thus π ≺ λΓ , i.e., λΓ weakly contains a non-0 finite-dimensional representation, which implies that Γ is amenable (see Dixmier [Di], 18.9.5 and 18.3.6), a contradiction. I do not know whether property (T) groups with the above properties exist. Finally, I also do not know if the following weakening of 13.2, (ii) is enough to guarantee amenability of Γ: Every action in FRERG(Γ, X, µ) has a dense conjugacy class in (ERG(Γ, X, µ), w). Remark. For each T ∈ ERG we consider the corresponding unitary operator UT0 ∈ U (L20 (X, µ)), induced by T on L20 (X, µ) and the associated set MT0 ⊆ {1, 2, . . . } ∪ {∞} of essential values of its multiplicity function, i.e., the set of numbers 1 ≤ n ≤ ∞ for which the multiplicity function of UT0 obtains the value n at a set of positive measure with respect to the maximal
13. TURBULENCE OF CONJUGACY IN THE ERGODIC ACTIONS
93
spectral type of UT0 . It is a major open problem in the spectral theory of ergodic transformations to characterize the sets MT0 corresponding to T ∈ ERG and it is indeed open whether any A ⊆ {1, 2, . . . } ∪ {∞} is of the form MT0 (see Katok-Thouvenot [KTh], 3.6). It is of course clear that the map ¯ ¯ = {1, 2, . . . } ∪ {∞} and we identify subsets T ∈ ERG 7→ MT0 ∈ 2N , where N ¯ of N with their characteristic functions, is Borel, so {MT0 : T ∈ ERG} is analytic but it is not clear whether it is Borel or not. An old problem, known as Rokhlin’s Problem, asked whether the sets {n}, n = 2, 3, 4, . . . are realized as MT0 for ergodic T . An affirmative answer was given recently by Ageev [A2] (and Ryzhikov [Ry2] for n = 2). Ageev’s method is to find for each n, a pair (Γn , γn ) of a countable group Γn and an element γn ∈ Γn such that for the generic a ∈ A(Γn , X, µ), if T = γna , then MT0 = {n}. Ageev suggests that there should be a spectral rigidity, i.e., for any pair (Γ, γ) of a countable group and γ ∈ Γ, Mγ0a is fixed for the generic a ∈ A(Γ, X, µ) (with respect to the weak topology w). He shows that sup(Mγ0a ) is indeed fixed for the generic a, provided (A(Γ, X, µ), w) has the Rokhlin property, i.e., admits a dense conjugacy class. Of course 10.7 asserts that this is indeed the case for all Γ, so that we can in fact infer (given that the map a 7→ Mγ0a is Borel and invariant under conjugacy, and using, e.g., Kechris [Kec2], 8.46) that we have full spectral rigidity, i.e., for any pair (Γ, γ), Mγ0a is fixed for the generic a ∈ A(Γ, X, µ). (A similar remark for sup(Mγoa ) is also made in Glasner-Thouvenot-Weiss [GTW]. A recent preprint of Ageev [A3] also contains this fact for Mγ0a .) Of course γ a will not be ergodic if a is not ergodic, so it may be more reasonable to look instead at the Polish space (ERG(Γ, X, µ), w). By 12.2, ERG(Γ, X, µ) is meager in (A(Γ, X, µ), w), if Γ has property (T), and dense Gδ if Γ does not have property (T), so it is irrelevant whether we use A(Γ, X, µ) or ERG(X, Γ, µ) if Γ does not have property (T). In any case, even if we work within ERG(Γ, X, µ), 13.1 implies that we still have a dense conjugacy class and thus Mγ0a is fixed for the generic a ∈ ERG(Γ, X, µ), for each infinite Γ. (B) We next study (generic) turbulence of the conjugacy action of the group Aut(X, µ) on ERG(Γ, X, µ). Theorem 13.3 (Kechris). Suppose a countable group Γ does not have property (T). Then the following are equivalent: (i) Every conjugacy class in ERG(Γ, X, µ) is meager in the topological space (ERG(Γ, X, µ), w). (ii) The conjugacy action of (Aut(X, µ), w) on (ERG(Γ, X, µ), w) is generically turbulent. Proof. It is of course enough to show that (i) ⇒ (ii). We work below in the weak topology. By 10.7, 12.2 (ii) and 12.3, there is an a ∈ ERG(Γ, X, µ) with dense conjugacy class which is not E0 -ergodic. It is enough to show that a is a turbulent point. Let E = Ea be the equivalence relation induced by a. Let
94
II. THE SPACE OF ACTIONS
for > 0, γ1 , . . . , γk ∈ Γ, A1 , . . . , A` ∈ MALGµ , U = {b ∈ ERG(Γ, X, µ) : ∀i ≤ k∀j ≤ `[µ(γib (Aj )∆γia (Aj )) < ]}. Fix also an open nbhd V of 1 ∈ Aut(X, µ). We will show that the local orbit O(a, U, V ) is dense in U . Since the conjugacy glass of a is dense in ERG(Γ, X, µ) it is enough to show that every T aT −1 ∈ U can be weakly approximated by elements of O(a, U, V ). Since E is ergodic, [E] is weakly dense in Aut(X, µ), so it is enough to take T aT −1 ∈ U with T ∈ [E]. Finally, as the periodic elements are uniformly dense in [E], we can assume that T is periodic. We can now proceed as in the proofs of 5.1, 5.2, using the fact that each γ a can be uniformly approximated by periodic elements of [E]. More precisely, let π : E → E0 be a Borel S homomorphism with the preimage of every E0 -class null. Let E0 = ∞ n=1 En , E1 ⊆ E2 ⊆ . . . finite Borel. Choose also δ > 0 sufficiently small (as needed below). Then choose periodic γ10 , . . . , γk0 ∈ [E], so that δu0 (γia , γi0 ) < δ/2, ∀i ≤ k, and, as in the proof of 5.2, choose N large enough so that there are T¯, γ¯1 , . . . , γ¯k ∈ [π −1 (EN ) ∩ E] with δu0 (T, T¯) < δ, δu0 (γi0 , γ¯i ) < δ/2, ∀i ≤ k. Then, again as in the proof of 5.2, we can find a continuous λ 7→ Xλ ∈ MALGµ , λ ∈ [0, 1], X0 = ∅, X1 = X, 0 ≤ λ ≤ λ0 ⇒ Xλ ⊆ Xλ0 , µ(Xλ \ Xλ0 ) ≤ λ0 − λ, and Xλ π −1 (EN )-invariant, so that it is also T¯, γ¯i (1 ≤ i ≤ k)-invariant. Put Tλ = T¯|Xλ ∪ id|(X \ Xλ ), so that Xλ is also Tλ , γ¯i (1 ≤ i ≤ k)invariant. Since T0 = 1, T1 = T¯, as in the proof of 5.1, we have that for A ∈ MALGµ , 1 ≤ i ≤ k, γi (A) ⊆ (T¯γ¯i T¯−1 (A)∆¯ γi (A)) ∩ Xλ . Tλ γ¯i T −1 (A)∆¯ λ
It follows that γi (A)) + δ µ(Tλ γia Tλ−1 (A)∆γia (A)) ≤ µ(Tλ γia Tλ−1 (A)∆¯ ≤ µ(Tλ γ¯i Tλ−1 (A)∆¯ γi (A)) + 2δ −1 ≤ µ(T¯γ¯i T¯ (A)∆¯ γi (A)) + 2δ ≤ µ(T γia T −1 (A)∆γia (A)) + 6δ, so if δ is small enough, clearly Tλ aTλ−1 ∈ U, ∀λ ∈ [0, 1]. Moreover, for any γ ∈ Γ, δu0 (T1 γ a T1−1 , T γ a T −1 ) = δu0 (T¯γ a T¯−1 , T γ a T −1 ) ≤ 2δ, and this completes the proof. 2 Foreman-Weiss [FW], using ideas about entropy of actions of amenable groups, show that every conjugacy class of an action of an amenable group is meager, thus we have the following result. Theorem 13.4 (Foreman-Weiss). If Γ is an infinite amenable group, the conjugacy action of the group (Aut(X, µ), w) on (ERG(Γ, X, µ), w) is generically turbulent. It is also clear that the conjugacy classes of actions of free groups are meager, so we have the next fact.
13. TURBULENCE OF CONJUGACY IN THE ERGODIC ACTIONS
95
Theorem 13.5 (Kechris). If Γ is a free group, then the conjugacy action of the group (Aut(X, µ), w) on (ERG(Γ, X, µ), w) is generically turbulent. In fact, more generally, if Γ, ∆ are countable groups without property (T) and the conjugacy action of (Aut(X, µ), w) on each one of the spaces (ERG(Γ, X, µ), w), (ERG(∆, X, µ), w) is generically turbulent, so in the action of (Aut(X, µ), w) on (ERG(Γ ∗ ∆, X, µ), w) (note that Γ ∗ ∆ does not have property (T) as well.) The following problem remains open. Problem 13.6. For which infinite countable groups Γ are the conjugacy classes in ERG(Γ, X, µ) meager in the weak topology of ERG(Γ, X, µ)? Is 13.3 true for property (T) groups? In the next section we will see an important fact concerning conjugacy classes in ERG(Γ, X, µ) for property (T) groups. (C) Although the question of turbulence for the conjugacy action of Aut(X, µ) on ERG(Γ, X, µ) is still unresolved for arbitrary Γ, one can still derive from 13.4 and work of Hjorth [Hj3] that the conjugacy equivalence relation on ERG(Γ, X, µ) cannot be classified by countable structures. Let U (H) be the unitary group of an infinite-dimensional separable Hilbert space H with the weak (equivalently the strong) topology. This is a Polish group. For each countable group Γ, let Rep(Γ, H) be the space of unitary representations of Γ on H, i.e., the closed subspace of H Γ consisting of all homomorphisms of Γ on H (see Appendix H). Finally, let Irr(Γ, H) be the subset of Rep(Γ, H) consisting of all irreducible representations. This is a Gδ subset of Rep(Γ, H), hence Polish. The group U (H) acts continuously on Rep(Γ, H) by conjugation. The corresponding equivalence relation is isomorphism or unitary equivalence of representations, in symbols π ∼ = ρ. We finally put π ≤ ρ if π is isomorphic to a subrepresentation of ρ. Hjorth [Hj3] has shown that there is a conjugacy invariant Gδ set GΓ,H ⊆ Irr(Γ, H) such that the action of U (H) on GΓ,H is turbulent, provided that Γ is not abelian-by-finite. This strengthens the result of Thoma [Th], as it implies that unitary equivalence of irreducible unitary representations cannot be classified by countable structures. (However Thoma’s result is used in Hjorth’s proof.) See Appendix H, (C) for an exposition of Hjorth’s work. It is known (see Appendix E) that one can assign in a Borel way to each π ∈ Rep(Γ, H) an action aπ ∈ A(Γ, X, µ), so that π∼ = ρ ⇒ aπ ∼ = aρ , and moreover if κa0π is the Koopman representation on L20 (X, µ) associated with aπ , then π ≤ κa0π . We now have the following result.
96
II. THE SPACE OF ACTIONS
Theorem 13.7 (Foreman-Weiss, Hjorth). Let Γ be an infinite countable group. Then conjugacy of measure preserving, free, ergodic actions of Γ cannot be classified by countable structures. Proof. If Γ is abelian-by-finite, this follows from 13.4. So assume Γ is not abelian-by-finite, and assume, towards a contradiction, that there is a countable language L, so that, denoting by XL the space of countable structures for L and by ∼ = the isomorphism relation on XL , there is a Borel function F : FRERG(Γ, X, µ) → XL with a∼ = b ⇔ F (a) ∼ = F (b). Now if π ∈ GΓ,H ⊆ Irr(Γ, H), aπ is weak mixing, by the remarks following E.1, and so if a0 is any fixed action in FRERG(Γ, X, µ), a0 × aπ is in FRERG(Γ, X, µ). Thus, for π, ρ ∈ GΓ,H ⊆ Irr(Γ, H), π∼ = ρ ⇒ F (a0 × aπ ) ∼ = F (a0 × aρ ). Since the conjugacy action on GΓ,H is turbulent, there is some M0 ∈ XL with F (a0 × aπ ) ∼ = M0 for comeager many π ∈ GΓ,H . So there is a ∈ FRERG(Γ, X, µ), for which on a comeager set of π ∈ GΓ,H , a0 × aπ ∼ = a. Since every conjugacy class in GΓ,H is meager in GΓ,H , there are uncountably many pairwise non-conjugate π ∈ Irr(Γ, H) such that a0 × aπ ∼ = a, and thus π ≤ κa0π ≤ κa00 ×aπ ∼ = κa0 , a which is a contradiction, as κ0 contains only countably many irreducible subrepresentations. 2 Remark. In a recent preprint, T¨ornquist [To2] has extended 13.7 to the set of measure preserving, free, ergodic actions of Γ that can be extended to free actions of ∆, for any (fixed) ∆ ≥ Γ. Remark. By choosing in the preceding proof a0 to be weak mixing, we see that 13.7 holds as well for measure preserving free, weak mixing actions of Γ. We also note that 13.7 and 2.2 of Hjorth-Kechris [HK1] imply that the equivalence relation E0 can be Borel reduced to conjugacy of such actions. It has been shown recently by Ioana-Kechris-Tsankov-Epstein (see [IKT]) that these facts are true as well for measure preserving, free, mixing actions, provided that the group Γ is not amenable. I do not know if this holds for all infinite Γ. (D) One can also consider unitary (spectral) equivalence in A(Γ, X, µ). Recall that two actions a ∈ A(Γ, X, µ), b ∈ A(Γ, Y, ν) are called unitarily (or spectrally) equivalent if the corresponding Koopman representations κa , κb are isomorphic. We then have the following analog of 13.7. Theorem 13.8 (Kechris). Let Γ be an infinite countable group. Then unitary equivalence of measure preserving, free, ergodic actions of Γ cannot be classified by countable structures.
13. TURBULENCE OF CONJUGACY IN THE ERGODIC ACTIONS
97
Proof. We will again consider cases depending on whether Γ is abelianby-finite or not. If Γ is not abelian-by-finite, then the argument in the proof of 13.7 works as well for unitary equivalence instead of conjugacy. Next consider the case where Γ is abelian-by-finite. Let ∆ Γ, ∆ infinite abelian, [Γ : ∆] < ∞. From now on we will use the notation and results of Appendices F, G. For a ∈ FRERG(∆, X, µ), let IndΓ∆ (a) ∈ FRERG(Γ, X × T, µ × νT ) be the induced action. Clearly a 7→ f (a) = IndΓ∆ (a) is Borel. Assume now, towards a contradiction, that there is a Borel map F : FRERG(Γ, X × T, µ × νT ) → XL with κc ∼ = κd ⇔ F (c) ∼ = F (d). Then for a, b ∈ FRERG(∆, X, µ), a∼ = b ⇒ f (a) ∼ = f (b) ⇒ F (f (a)) ∼ = F (f (b)). So, by 13.4 and 10.8, there is a comeager set A ⊆ FRERG(∆, X, µ) such that for a, b ∈ A, F (f (a)) ∼ = F (f (b)) and thus Γ Γ κInd∆ (a) ∼ = κInd∆ (b) ,
i.e., Γ Γ κInd∆ (a) ∼ = IndΓ∆ (κa ) ∼ = IndΓ∆ (κb ) ∼ = κInd∆ (b) ∼ IndΓ (κb )|∆. So if σa , σb are the maximal specand therefore IndΓ∆ (κa )|∆ = ∆ P P tral types for κa , κb , resp., so that t∈T (θˆt )∗ σa , t∈T (θˆt )∗ σb are the maximal spectral types for IndΓ∆ (κa )|∆, IndΓ∆ (κb )|∆, resp., then we have
(1)
a, b ∈ A ⇒
X t∈T
(θˆt )∗ σa ∼
X
(θˆt )∗ σb .
t∈T
We will next use some ideas from Choksi-Nadkarni [CN] (see also [Na, ˆ = the space of probabilCh. 8]). First recall that for each ρ ∈ P (∆) ˆ ˆ ˆ (see, e.g., ity measures on ∆, {ν ∈ P (∆) : ν ⊥ ρ} is a Gδ set in P (∆) a Kechris-Sofronidis [KS]). Since a ∈ A(∆, X, µ) 7→ κ0 ∈ Rep(∆, L20 (X, µ)) is continuous, so that a 7→ σκa0 is also continuous, we have that {a ∈ P ˆ a A(∆, X, µ) : σκa0 ⊥ ρ} and Bρ = {a ∈ A(∆, X, µ) : t∈T (θt )∗ σκ0 ⊥ ρ} are Gδ as well (we work here in the weak topology of A(∆, X, µ)) and also conjugacy invariant. We will find a free ergodic action c0 ∈ A(∆, X, µ) such that the maximal spectral type σκc0 of κc00 is supported by a countable set, 0 P P and thus so is (θˆt )∗ σ c0 , and, as a consequence, (θˆt )∗ σ c0 ⊥ λ, t∈T
t∈T
κ0
κ0
ˆ where λ = η∆ ˆ is the Haar measure on ∆. Since the conjugacy class of c0 in A(∆, X, µ) is dense, by 13.2, it follows that Bλ is comeager and thus there P is b0 ∈ Bλ ∩ A, i.e., we have b0 ∈ A, t∈T (θˆt )∗ σκb0 = σ0 ⊥ λ. Let now s 0
be the Gaussian shift on R∆ associated to ϕ1 = the characteristic function
98
II. THE SPACE OF ACTIONS
of {1} ⊆ ∆ (see Appendix D). Then (see Appendix F, (C)) the maximal P P spectral type for κs0 is λ and t∈T (θˆt )∗ σκs0 ∼ t∈T (θˆt )∗ λ = |T |λ (as each θˆt preserves λ). Thus s ∈ Bσ0 . Again the conjugacy class of s is dense in A(∆, X, µ), as s is free and ergodic, so Bσ0 is comeager. Thus there is P P a0 ∈ A ∩ Bσ0 , i.e., a0 ∈ A and t∈T (θˆt )∗ σκa0 ⊥ t∈T (θˆt )∗ σκb0 and there0 0 P P fore t∈T (θˆt )∗ σa0 6∼ t∈T (θˆt )∗ σb0 , where σa0 = δ1 + σκa0 is the maximal 0 spectral type for κa0 and similarly for σb0 (where δ1 is the Dirac measure at ˆ This contradicts (1). 1 ∈ ∆). Construction of (an isomorphic copy) of c0 . By the next to last paragraph of Section 9, we can assume that ∆ is a dense subgroup of a compact Polish group G. We take c0 to be the translation action of ∆ on (G, ηG ), where ηG is the Haar measure on G. Clearly c0 is free and ergodic. Then, ˆ is an orthonormal basis for L2 (G, ηG ) and using the fact that {χ : χ ∈ G} Cχ is invariant under the translation action, it is easy to check that the maximal spectral type for κc00 is supported by a countable set and the proof is complete. 2 Finally we have the following result. Theorem 13.9. Let Γ be an infinite countable group. Then weak isomorphism of measure preserving, free, ergodic actions of Γ cannot be classified by countable structures. Proof. If Γ is not abelian-by-finite, the argument in the proof of 13.7 works as well. If Γ is abelian-by-finite, then the argument in the proof of 13.8 applies and the proof is complete. 2 Remark. Again 13.8 and 13.9 hold as well for weak mixing actions. It appears that weak isomorphism ∼ =w is of a different nature than isomorphism ∼ = or unitary equivalence, which are both Borel reducible to equivalence relations induced by continuous actions of Polish groups. For example, the following is open. Problem 13.10. Let Γ be a countable infinite group. Let E1 be the following equivalence relation on RN : (xn )E1 (yn ) ⇔ ∃n∀m ≥ n(xm = ym ). ∼w on FRERG(Γ, X, µ)? Can E1 be Borel reduced to = An affirmative answer would imply that ∼ =w cannot be Borel reducible to equivalence relations induced by continuous actions of Polish groups. 14. Conjugacy in ergodic actions of property (T) groups (A) We will now use a technique employed in Hjorth [Hj4] to study conjugacy classes of ergodic actions of property (T) groups. It traces its origins in separability arguments used in the context of operator algebras
14. CONJUGACY IN ERGODIC ACTIONS OF PROPERTY (T) GROUPS
99
in Popa [Po1]. In essence, Hjorth’s proof, from the point of view presented here, showed that the conjugacy classes in ERG(Γ, X, µ) are clopen in (ERG(Γ, X, µ), u). However a more careful analysis, which is encapsulated in Lemma 14.1 below, can be used to derive stronger conclusions, see Theorem 14.2. Let Γ = {γn } be a countable group and let X δΓ,u (a, b) = 2−n δu (γna , γnb ) n
be the metric giving the uniform topology on A(Γ, X, µ). Also if E is a countable measure preserving equivalence relation on (X, µ), then letting ϕ vary over all partial Borel automorphisms ϕ : A → B, where A, B are Borel sets, we put [[E]] = {ϕ : ϕ(x)Ex, ∀x ∈ A}. Note that if ϕ ∈ [[E]], then ϕ is measure preserving on its domain. Below for two equivalence relations E, F , we define their join, E ∨ F , to be the smallest equivalence relation containing E and F . Lemma 14.1. Let Γ be a countable group with property (T). Then given δ > 0, there is η > 0, such that for any a, b ∈ A(Γ, X, µ), if δΓ,u (a, b) < η, then there is ϕ : A → B, ϕ ∈ [[Ea ∨ Eb ]], such that A is a-invariant, B is b-invariant, ϕ(a|A)ϕ−1 = b|B, and µ(A) > 1 − 16δ 2 . Moreover, µ({x ∈ A : ϕ(x) 6= x}) ≤ 4δ 2 . Proof. Put below E = Ea ∨Eb . The main idea, which comes from Hjorth [Hj4], is to consider the product action a×b on X 2 . Notice that it leaves E ⊆ X 2 invariant andR the σ-finite measure M on E invariant, where for Borel A ⊆ E, M (A) = card(Ax )dµ(x). Thus it gives a unitary representation of Γ on L2 (E, M ), which, viewed as an action of Γ on L2 (E, M ), is defined by γ · f (x, y) = f ((γ −1 )a (x), (γ −1 )b (y)). Claim. If ∆ : E → C is the diagonal function, ∆(x, y) = 1 if x = y, ∆(x, y) = 0, if x 6= y, then kγ −1 · ∆ − ∆k22 = 2δu (γ a , γ b ) Proof of the claim. We have R kγ −1 · ∆ − ∆k22 = RE |∆(γ a (x), γ b (y)) − ∆(x, y)|2 dM (x, y) = x=y |∆(γ a (x), γ b (y)) − 1|2 dM (x, y) R + x6=y,(x,y)∈E ∆(γ a (x), γ b (y))dM (x, y) = µ({x : γ a (x) 6= γ b (x)}) +M ({(x, y) ∈ E : x 6= y & γ a (x) = γ b (y)}).
100
II. THE SPACE OF ACTIONS
Now M ({(x, y) ∈ E : x 6= y & γ a (x) = γ b (y)}) = M ({(x, y) ∈ E : x 6= y & (γ −1 )b γ a (x) = y}) = µ({x : (γ −1 )b γ a (x) 6= x}) = µ({x : γ a (x) 6= γ b (x)}) = δu (γ a , γ b ), so kγ −1 · ∆ − ∆k22 = 2δu (γ a , γ b ). Recall now that Γ admits a Kazhdan pair (Q, ), i.e., a pair consisting of a finite subset Q ⊆ Γ and > 0 such that for any unitary representation π : Γ → U (H), if there is v ∈ H which is (Q, )-invariant, i.e., kπ(γ)(v)−vk < kvk, ∀γ ∈ Q, then there is v 6= 0 which is Γ-invariant. Recall also (see, e.g., Bekka-de la Harpe-Valette [BdlHV]) that if (Q, ) is a Kazhdan pair for Γ and δ > 0, then for any unitary representation π : Γ → U (H), if v ∈ H is (Q, δ)-invariant, then there is Γ-invariant v1 such that kv − v1 k ≤ δkvk. So given δ > 0, choose η > 0 small enough such that if δΓ,u (a, b) < η, then 2δu ((γ −1 )a , (γ −1 )b ) < δ 2 2 , ∀γ ∈ Q. Then, applying the above to the representation of Γ on L2 (E, M ) and the vector ∆, we can find f ∈ L2 (E, M ) with k∆ − f k2 ≤ δ which is Γ-invariant, i.e., f (γ a (x), γ b (y)) = f (x, y), ∀γ ∈ Γ. Following Hjorth [Hj4] put R(x, y) ⇔ xEy & |f (x, y) − 1| < 1/2. Then clearly R is a × b-invariant, i.e., R(x, y) ⇔ R(γ a (x), γ b (y)). Let C = {x : There is unique y with R(x, y)}, and for x ∈ C, put T (x) = the unique y with R(x, y). Let also D = T (C). Clearly C is a-invariant, D is b-invariant and T (γ a (x)) = γ b (T (x)), for x ∈ C. Let now B = {y ∈ D : There is unique x ∈ C with R(x, y)}, Then
T −1
is well-defined on B and put A = T −1 (B) ⊆ C.
Clearly A is a-invariant, B is b-invariant and if ϕ = T |A, then ϕ ∈ [[E]] and ϕ(a|A)ϕ−1 = b|B. It remains to prove that µ(A) (= µ(B)) ≥ 1 − 16δ 2 . Claim. µ(D) ≥ 1 − 12δ 2 .
14. CONJUGACY IN ERGODIC ACTIONS OF PROPERTY (T) GROUPS
101
Proof of the claim. Let D1 = {y : ∀x(x, y) 6∈ R}, D2 = {x : ∃z 6= xR(x, z)}, D3 = {y : ∃x 6= yR(x, y)}. We claim that X \ D ⊆ D1 ∪ D2 ∪ D3 . Indeed, let y ∈ / D and y ∈ / D1 . Then there is x with R(x, y), thus x ∈ / C, so there is z 6= y with R(x, z). If x = y, then clearly y ∈ D2 , while if x 6= y, then y ∈ D3 . R R P Recall now that for any Borel h ≥ 0, hdM = ( x∈[y]E h(x, y))dµ(y). Thus RP δ 2 ≥ kf − ∆k22 = |f (x, y) − ∆(x, y)|2 dµ(y) R 1x∈[y]E 1) ≥ D1 4 dµ(y) = µ(D 4 , so µ(D1 ) ≤ 4δ 2 . Also δ 2 ≥ kf − ∆k22 ≥
Z D3
1 µ(D3 ) dµ(y) = , 4 4
4δ 2 .
so µ(D3 ) ≤ R R P Since also for Borel h ≥ 0, hdM = ( y∈[x]E h(x, y))dµ(x), and D2 = {x : ∃y 6= xR(x, y)}, we have RP |f (x, y) − ∆(x, y)|2 dµ(x) δ 2 ≥ kf − ∆k22 = R 1y∈[x]E µ(D2 ) ≥ D2 4 dµ = 4 , thus µ(D2 ) ≤ 4δ 2 , and finally µ(X \ D) ≤ 12δ 2 and µ(D) ≥ 1 − 12δ 2 . Let now F = {y : ∃x1 6= x2 (x1 , x2 ∈ C & R(x1 , y) & R(x2 , y))}. Again F is b-invariant and as before δ 2 ≥ kf − ∆k22 ≥
Z F
µ(F ) 1 dµ(y) = , 4 4
4δ 2 .
so µ(F ) ≤ Clearly B = D\F and thus µ(B) ≥ 1−12δ 2 −4δ 2 = 1−16δ 2 . For the final assertion of the lemma, note that {x ∈ A : ϕ(x) 6= x} ⊆ D2 . 2 We now have the following consequence. Theorem 14.2. Let Γ be a countable group with property (T). Then (i) ERG(Γ, X, µ) is clopen in (A(Γ, X, µ), u), (ii) Every conjugacy class in ERG(Γ, X, µ) is clopen in (A(Γ, X, µ), u), (iii) WMIX(Γ, X, µ) is clopen in (A(Γ, X, µ), u). Proof. Let a ∈ ERG(Γ, X, µ) and let η be chosen as in Lemma 14.1 for δ < 1/4. Let then δΓ,u (a, b) < η. Then for ϕ, A, B as in that lemma, µ(A) > 0, so by the ergodicity of a, µ(A) = µ(B) = 1 and so ϕ ∈ Aut(X, µ) and ϕaϕ−1 = b, thus b is conjugate to a and b is ergodic. It follows that ERG(Γ, X, µ) is open in u and so is every conjugacy class in ERG(Γ, X, µ).
102
II. THE SPACE OF ACTIONS
Now ERG(Γ, X, µ) is also closed in u, by 12.2 i), and thus it is clopen in u and therefore the same is true for each conjugacy class contained in it. That WMIX(Γ, X, µ) is clopen in A(Γ, X, µ) follows from the fact that a ∈ WMIX(Γ, X, µ) iff a × a ∈ ERG(Γ, X × X, µ × µ) and the map a 7→ a × a is uniformly continuous. 2 A similar argument shows that for groups Γ with property (T), the set FED(Γ, X, µ) is open and each conjugacy class contained in it is open in (A(Γ, X, µ), u). It is clear from 14.2, and the separability of each full group in the uniform topology, that we now have the following. Theorem 14.3 (Hjorth). Let Γ have property (T). Then every orbit equivalence class in ERG(Γ, X, µ) contains only countably many conjugacy classes. Hjorth derives from this a result about the complexity of classification of ergodic actions of Γ up to orbit equivalence. Theorem 14.4 (Hjorth). Let Γ be a countable infinite group with property (T). Then orbit equivalence of measure preserving, free, ergodic actions on Γ cannot be classified by countable structures. Proof. Using the notation and the argument in the proof of 13.7, and noting that, as Γ has property (T), Γ is not abelian-by-finite, we conclude, if the theorem fails, that there is a ∈ FRERG(Γ, X, µ) such that we have (a0 × aπ )OEa for comeager many π ∈ GΓ,H . Then, by 14.3, there is {an } ⊆ FRERG(Γ, X, µ) such that on a comeager set of π ∈ GΓ,H , a0 × aπ ∼ = an for some n. So there is b ∈ FRERG(Γ, X, µ) such that for a non-meager set of π ∈ GΓ,H , a0 × aπ ∼ 2 = b, from which we have a contradiction as in 13.7. Remark. I do not know what countable groups Γ have the property that ERG(Γ, X, µ) is closed in u and similarly what groups Γ have the property that ERG(Γ, X, µ) is open in u. It may very well be that the last property characterizes the property (T) groups. However, it is easy to see that if N1 , resp., N2 , denote the classes of groups Γ for which ERG(Γ, X, µ) is not uniformly closed, resp., not uniformly open, then Z ∈ Ni , and if Γ ∈ Ni and Γ is a factor of ∆, then ∆ ∈ Ni , for i = 1, 2. For example, for the free group Fn (n ≥ 1), ERG(Fn , X, µ) is neither closed nor open in u. In fact, it is not hard to see that ERG(Fn , X, µ) is not even Fσ in u (it is of course Gδ in u). Indeed, fixing free generators γ1 , . . . , γn for Fn , consider the closed in u set C of all actions a such that γia is aperiodic for all i ≤ n. Then clearly the ergodic actions are uniformly dense in C by 2.4. But also the non-ergodic actions are uniformly dense in C (which, be the Baire Category Theorem implies that the ergodic actions cannot be Fσ in u). There are several ways to see this but perhaps the easiest, pointed out by Tsankov, is to use the proof of 2.8 and notice that the set A is invariant under the induced transformation TA .
14. CONJUGACY IN ERGODIC ACTIONS OF PROPERTY (T) GROUPS
103
Remark. There is a version of 14.1 valid for an arbitrary countable group Γ. Consider the space A(Γ, X, µ) and define the metric δΓ,∞ (a, b) = sup δu (γ a , γ b ). γ∈Γ
Clearly δΓ,∞ ≥ δΓ,u and the topology induced by δΓ,∞ contains the uniform topology. It is also easy to check that δΓ,∞ is a complete metric on A(Γ, X, µ). Given a, b ∈ A(Γ, X, µ) such that δΓ,∞ (a, b) < δ 2 /2 and going through the proof of 14.1 we see that kγ · ∆ − ∆k2 < δ, ∀γ ∈ Γ. Thus Γ · ∆ is a Γinvariant set, contained in the ball of radius δ around ∆. Let f be the unique element of least norm in the closed convex hull of Γ·∆. Then f is Γ-invariant and k∆ − f k ≤ δ. The rest of the proof of 14.1 can be repeated verbatim. Thus we have that for any countable group Γ and a, b ∈ A(Γ, X, µ), if δΓ,∞ (a, b) < δ 2 /2 then there is ϕ : A → B, ϕ ∈ [[Ea ∨ Eb ]] such that A is a-invariant, B is b-invariant, ϕ(a|A)ϕ−1 = b|B and µ(A) > 1 − 16δ 2 , µ({x ∈ A : ϕ(x) 6= x}) ≤ 4δ 2 . It follows, as in 14.2, that for any countable group Γ, ERG(Γ, X, µ) is clopen in the δΓ,∞ -topology and so is every conjugacy class in ERG(Γ, X, µ). Similarly WMIX(Γ, X, µ) is clopen in this topology. Finally, we note that if Γ has property (T), then the uniform topology coincides with the δΓ,∞ -topology. For this it is enough to show that for any δ > 0, there is η > 0 such that if δΓ,u (a, b) < η, then δΓ,∞ (a, b) ≤ 56δ 2 . Fix δ > 0 and let η be given by 14.1. If δΓ,u (a, b) < η find ϕ : A → B as in Lemma 14.1 and extend ϕ to T ∈ Aut(X, µ), so that T (X \A) = X \B. Now for each γ ∈ Γ, T γ a T −1 (x) = γ b (x), for each x ∈ B, thus δu (γ b , T γ a T −1 ) ≤ µ(X \ B) < 16δ 2 and δu (T, 1) ≤ µ({x ∈ A : ϕ(x) 6= x} ∪ (X \ A)) < 20δ 2 . Thus δu (γ b , γ a ) ≤ δu (γ b , T γ a T −1 ) + δu (T γ a T −1 , γ a ) ≤ 16δ 2 + 2δu (T, 1) ≤ 56δ 2 , so δΓ,∞ (a, b) ≤ 56δ 2 . (B) Lemma 14.1 also has implications concerning the outer automorphism group of equivalence relations induced by actions of property (T) groups. Theorem 14.5 (Gefter-Golodets [GG]). Suppose the countable group Γ has property (T) and let a ∈ ERG(Γ, X, µ). Denote by Ca ⊆ N [Ea ] the centralizer of a, i.e., the set of all T ∈ Aut(X, µ) that preserve the action: T aT −1 = a. Then Ca [Ea ] = {T S : T ∈ Ca , S ∈ [Ea ]} is clopen in N [Ea ]. In particular, if Ca = {1}, then Out(Ea ) is countable. Proof. Put E = Ea , C = Ca . For T ∈ N [E], let X ρ(T ) = 2−n δu (γna , T γna T −1 ), where Γ = {γn }. Then it is enough to show that there is some η > 0 such that for T ∈ N [E], ρ(T ) < η ⇒ T ∈ C[E].
104
II. THE SPACE OF ACTIONS
Put b = T aT −1 . Then Eb ⊆ E. Moreover ρ(T ) = δΓ,u (a, b). So if η is as in 14.1 for δ < 41 and ρ(T ) < η, then (as a, b are ergodic and Ea ∨ Eb = E), there is S ∈ [E] with SaS −1 = b = T aT −1 , and thus we have (S −1 T )a(S −1 T )−1 = a, i.e., S −1 T ∈ C, so T ∈ SC ⊆ [E]C = C[E]. 2 Note also the following result. Theorem 14.6. Let Γ be a countable group and let a ∈ FRERG(Γ, X, µ). Then Ca ∩ [Ea ] = {1} if one of the two conditions below hold: (i) (Gefter [Ge]) Γ is ICC. (ii) Γ is centerless and every infinite subgroup of Γ acts ergodically (e.g., the action is mixing). S S Proof. Suppose T ∈ [Ea ] is in Ca . Let T = i γi |Ai , where A = i Ai is a partition of X into Borel sets of positive measure. By ergodicity Ai meets every Ea -class. We will show that each γi = 1. For xEa y, let α(x, y) · x = y be the corresponding cocycle α : Ea → Γ. For each Ea -class C, let ΓC i = {α(x, y) : x, y ∈ C ∩ Ai } ⊆ Γ. Then by ergodicity ΓC i = Γi is constant for (almost) all C. Claim. If δ ∈ Γi , then γi δ = δγi . Proof of the claim. Indeed, if δ = α(x, y), where x, y ∈ Ai , xEa y, we have δ · x = y and γi δ · x = γi · y = T (y) = T (δ · x) = δ · T (x) = δγi · x. Since the action is free, γi δ = δγi . Thus γi commutes with every element of hΓi i. Consider first case (ii). Since Ai meets every Ea -class infinitely often, Γi is infinite and so is hΓi i, thus hΓi i acts ergodically. Now hΓi i · Ai is conull and thus contains a conull Ea -invariant set, thus for (almost) all C, C is a hΓi i-orbit, thus (again by the freeness of the action) hΓi i = Γ. So γi = 1. In case (i), notice that for the conjugacy action of Γ on itself the stabilizer of γi contains hΓi i. If γi 6= 1, the conjugacy class of γi is infinite, so hΓi i has infinite index in Γ, which implies that there are infinitely many {δn } ⊆ Γ with {δn hΓi i · Ai } pairwise disjoint, a contradiction. (There is actually a simpler proof in case (i); see Golodets [Gol] or Furman [Fu1]: Suppose T ∈ Ca ∩ [Ea ]. There is Borel f (x) = γx ∈ Γ with T (x) = γx · x. Then γδ·x = δγx δ −1 . Thus f∗ µ is a probability measure on Γ which in conjugacy-invariant, so, since Γ is ICC, concentrates on 1, i.e., f (x) = 1 and so T = 1.) 2 (C) Finally, one can use these ideas to give a proof of the result of Ozawa [O] that no full group [F ] contains (algebraically) every countable
15. CONNECTEDNESS IN THE SPACE OF ACTIONS
105
group. Assume such F existed, towards a contradiction, and without loss of generality take F to be ergodic (by enlarging it if necessary). Let us say that a non-trivial countable group Γ ≤ [F ] is nice if {x : γ(x) = x, ∀γ ∈ Γ} has measure < 12 . Notice that if Γ ≤ [F ], then there is an isomorphic copy Γ0 of Γ, Γ0 ≤ [F ], such that Γ0 is nice. Indeed, if Γ is not nice itself, let C = {x : γ(x) = x, ∀γ ∈ Γ}. Then 1 > µ(C) ≥ 12 , so if D = X \ C, 0 < µ(D) ≤ 12 . Note that, by the ergodicity of F , both C, D are complete sections of F . Let C1 , . . . ,SCn be pairwise disjoint subsets of C such that µ(Ci ) = µ(D) and µ(C \ ni=1 Ci ) < 12 . Let ϕi : D → Ci be a Borel 1-1 correspondence with ϕi (x)F x, ∀x ∈ D. Notice that C, D and each Ci are of course Γinvariant. For each γ ∈ Γ, let γ 0 ∈ [F ] be defined as follows: γ 0 (x) = γ(x), Sn 0 if x ∈ D; γ (x) = γ(x) = x, if x ∈ C \ i=1 Ci ; γ 0 (x) = ϕi (γ(ϕ−1 i (x)), if x ∈ Ci . Clearly γ 7→ γ 0 S is an isomorphism of Γ with Γ0 ≤ [F ] and {x : γ 0 (x) = x, ∀γ ∈ Γ} = C \ ni=1 Ci , so Γ0 is nice. As in Ozawa [O], we fix a countable property (T) group Γ which has uncountably many pairwise non-isomorphic factors {Γi }i∈I with each Γi simple. Fix epimorphisms πi : Γ → Γi . Then the kernels Ni = ker(πi ), i ∈ I, are all distinct. Assume, towards a contradiction, that each Γi can be (algebraically) embedded into [F ]. By the preceding remarks, we can assume that the image of each embedding is nice, i.e., there is an isomorphism ϕi : Γi → [F ] with ϕi (Γi ) nice. Consider then the action ai of Γ on (X, µ) given by γ ai = ϕi (πi (γ)). Thus for all i, µ({x : γ ai (x) = x, ∀γ ∈ Γ}) < 12 . Clearly {a ∈ A(Γ, X, µ) : γ a ∈ [F ], ∀γ ∈ Γ} is separable in the uniform topology, so there are i 6= j with δΓ,u (ai , aj ) < η, where η is chosen small enough so that, by Lemma 14.1, there is ϕ : A → B, ϕ ∈ [[F ]] with A ai -invariant, B aj -invariant, ϕ(ai |A)ϕ−1 = (aj |B), and µ(A) > 21 . Since Ni 6= Nj , suppose, without loss of generality, that Nj 6⊆ Ni . Then πi (Nj ) is a non-trivial normal subgroup of Γi , so πi (Nj ) = Γi . But if γ ∈ Nj , γ aj |B = id, so γ ai |A = id, thus A ⊆ {x : γ ai (x) = x, ∀γ ∈ Γ}, which is a contradiction, as µ(A) > 12 . 15. Connectedness in the space of actions (A) Recall from 2.8 that the group Aut(X, µ), which we can identify with the space of Z-actions, A(Z, X, µ), is contractible in the uniform topology and the same argument shows that each open ball in δu around the identity is also contractible, so that in particular (Aut(X, µ), u) is locally path connected. By 2.9, (Aut(X, µ), w) is homeomorphic to `2 , so again it is locally path connected. We next prove another local connectedness fact about Aut(X, µ) which can serve as a prototype for similar results for other A(Γ, X, µ) that we will discuss later on. We start with a lemma that can be proved by using the argument for turbulence in 5.1. Below for T ∈ Aut(X, µ), > 0, we put VT, = {S ∈ Aut(X, µ) : δw (S, T ) < },
106
II. THE SPACE OF ACTIONS
where δw is the metric giving the weak topology on Aut(X, µ), as in Section 1, (B). Lemma 15.1. Let T ∈ Aut(X, µ), T ∈ [E], with E hyperfinite, g ∈ [E], δ, > 0 and gT g −1 ∈ VT, . Then there is a continuous path in ([E], u), λ 7→ gλ , λ ∈ [0, 1], with g0 = 1 and gλ T gλ−1 ∈ VT, , for which we have δu0 (gT g −1 , g1 T g1−1 ) < δ. Similarly if VT, is replaced by the open ball of radius around T in the metric δu0 . Using this one can prove the following fact, due to Rosendal and the author. Lemma 15.2. Let T ∈ APER, T ∈ [E], with E ergodic, hyperfinite, and let S ∈ VT, . Then there is a continuous map λ 7→ hλ , λ ∈ [0, 1), in ([E], u) with −1 h0 = 1, hλ T h−1 λ ∈ VT, and hλ T hλ → S weakly as λ → 1. In particular, there is a continuous path in (VT, , w) connecting T to S. Proof. Let V = VT, . Fix a sequence n ↓ 0 such that VS,2n ⊆ V , ∀n. By the aperiodicity of T and the ergodicity of E, {gT g −1 : g ∈ [E]} is weakly dense in Aut(X, µ), so we can find g ∈ [E] such that gT g −1 is as close to S as we want, in the weak topology, so by 15.1 we can find a continuous path h1,λ −1 in ([E], u) such that h1,0 = 1, h1,λ T h−1 1,λ ∈ V , T1 = h1,1 T h1,1 ∈ VS,1 , and so S ∈ VT1 ,1 , thus VT1 ,1 ⊆ VS,21 ⊆ V . By the same argument, we can find a continuous path h2,λ in ([E], u) such that h2,0 = 1, h2,λ T1 h−1 2,λ ∈ VT1 ,1 , T2 = −1 h2,1 T1 h2,1 ∈ VS,2 , and so S ∈ VT2 ,2 , thus VT2 ,2 ⊆ VS,22 ⊆ V , . . . Fix finally α1 < α2 < · · · → 1. Then define hλ as follows: hλ = h
λ 1, a
, λ ∈ [0, a1 ],
1
∈ [a1 , a2 ],
hλ = h
(λ−a1 ) h1,1 , λ 2, (a −a ) 2 1
hλ = h
(λ−a3 ) h2,1 h1,1 , λ 3, (a −a ) 3 2
∈ [a2 , a3 ], . . .
It is easy to see that this works.
2
Theorem 15.3. Let C ⊆ Aut(X, µ) contain the conjugacy class of an aperiodic T ∈ Aut(X, µ). Then (C, w) is path connected and locally path connected. Proof. From 15.2, by taking > 1, and using also 5.4, we see that C is path connected. Also by considering the family of all VT, ∩ C with T ∈ APER ∩ C, we see that it is locally path connected. 2 A similar argument, using now the proof of 5.2, shows the following. Theorem 15.4. Let E be an ergodic equivalence relation which is not E0 ergodic. Let C ⊆ APER ∩ [E] contain the conjugacy class in [E] of an aperiodic T ∈ [E]. Then (C, u) is path connected and locally path connected. Moreover, a continuous path from T to any S ∈ C can be found that has
15. CONNECTEDNESS IN THE SPACE OF ACTIONS
107
the form Tλ = gλ T gλ−1 , λ ∈ [0, 1), T1 = S, where λ 7→ gλ , λ ∈ [0, 1), is continuous in ([E], u). Corollary 15.5. Both (APER, u) and (ERG, u) are path connected. Proof. Let S, T ∈ ERG. By Dye’s Theorem 3.13, there is U ∈ Aut(X, µ) such that U [ES ]U −1 = [ET ] and thus U SU −1 ∈ [ET ]. Since U SU −1 , T ∈ [ET ], it follows from 15.4 that there is a continuous path in (ERG ∩ [ET ], u) from U SU −1 to T . By 2.8, there is a continuous path Uλ , 0 ≤ λ ≤ 1, in (Aut(X, µ), u) from 1 to U . Clearly Uλ SUλ−1 is a continuous path in (ERG, u) from S to U SU −1 , so there is a continuous path in (ERG, u) from S to T . The argument for APER is similar (using the fact, see 5.4, that for any S, T ∈ APER, there are S0 , T0 ∈ ERG with ES ⊆ ES0 , E ⊆ ET0 ). As pointed out by Tsankov, it can be also proved directly as in 2.8. 2 (B) We now consider a countable group Γ and the space A(Γ, X, µ). First let us note the following fact. Proposition 15.6. For any countable group Γ, the conjugacy class of any a ∈ A(Γ, X, µ) is path connected in (A(Γ, X, µ), u) (and thus also in the space (A(Γ, X, µ), w)). Proof. Let a ∈ A(Γ, X, µ) and T ∈ Aut(X, µ). By 2.8, there is a continuous path t 7→ Tt in (Aut(X, µ), u) with T0 = 1, T1 = T . Let at = Tt aTt−1 . Then t 7→ at is a continuous path in (A(Γ, X, µ), u) with a0 = a, a1 = T aT −1 . 2 Corollary 15.7 (a weak version of 15.11). For any countable group Γ, the space (A(Γ, X, µ), w) is connected. Proof. From 10.7 and 15.6.
2
This fact can be reformulated as follows. For any topological group G, a variety on G is a closed subset on GN , where N = 1, 2, . . . , N, of the form {(g1 , g2 , . . . ) ∈ GN : ∀i ∈ I(wig1 ,g2 ... = 1)}, where {wi }i∈I is a family of words in the free group FN . Recalling the remark in Section 10, (A), we can rephrase 15.7 as follows: Every variety in (Aut(X, µ), w) is connected. The next question is whether (A(Γ, X, µ), w) is path connected. The answer turns out to be positive. The main fact is that it holds for groups that do not have property (T). In fact, for such groups we have a complete analog of 15.3. Theorem 15.8 (Kechris). Let Γ be a countable group that does not satisfy property (T). Let C ⊆ A(Γ, X, µ) contain the conjugacy class of an a ∈ A(Γ, X, µ), which is ergodic, not E0 -ergodic and has a weakly dense conjugacy class. Then (C, w) is path connected and locally path connected. In particular, (A(Γ, X, µ), w) has these properties.
108
II. THE SPACE OF ACTIONS
~ = {B1 , . . . , Bk } ⊆ MALGµ , > Proof. For ~γ = {γ1 , . . . , γ` } ⊆ Γ, B 0, a ∈ A(Γ, X, µ), let Ua,,~γ ,B~ = {b ∈ A(Γ, X, µ) : ∀j ≤ `, ∀i ≤ k(µ(γja (Bi )∆γjb (Bi )) < )}. 0 Let also δΓ,u on A(Γ, X, µ) be defined by X 0 δΓ,u (a, b) = 2−n δu0 (γna , γnb ), 0 for some fixed enumeration Γ = {γn }, so that δΓ,u is a complete compatible metric in u. First we have an analog of 15.1, whose proof is as in 13.3.
Lemma 15.9. Let a ∈ A(Γ, X, µ) be ergodic but not E0 -ergodic, T ∈ [Ea ], δ > 0 and T aT −1 ∈ Ua,,~γ ,A~ . Then there is a continuous path in ([Ea ], u), λ 7→ Tλ , λ ∈ [0, 1], with T0 = 1, Tλ aTλ−1 ∈ Ua,,~γ ,A~ and 0 δΓ,u (T1 aT1−1 , T aT −1 ) < δ.
From this, as in 15.2, we can deduce the following lemma. Lemma 15.10. Let a ∈ ERG(Γ, X, µ) \ E0 RG(Γ, X, µ) have weakly dense conjugacy class in A(Γ, X, µ). Let b ∈ Ua,,~γ ,A~ . Then there is continuous λ 7→ Tλ , λ ∈ [0, 1), in (Aut(X, µ), u) with T0 = 1, Tλ aTλ−1 ∈ Ua,,~γ ,A~ and Tλ aTλ−1 → b weakly as λ → 1. In particular, there is a continuous path in (Ua,,~γ ,A~ , w) connecting a to b. The theorem is then immediate from 15.10. The last assertion also follows using 10.7, 12.4 ii). 2 Theorem 15.11 (Kechris). Let Γ be a countable group. Then the space (A(Γ, X, µ), w) is path connected and locally path connected. Proof. Let ∆ = Γ ∗ Z, so that ∆ does not have property (T). Then by Section 10, (G), A(∆, X, µ) is homeomorphic to (A(Γ, X, µ) × A(Z, X, µ), and since, by 15.8, A(∆, X, µ) is path connected and locally path connected, so is A(Γ, X, µ). 2 Again this can be reformulated as: Every variety in (Aut(X, µ), w) is path connected and locally path connected. It follows from the proofs of 15.8 and 15.11 that for any Γ, the space (FR(Γ, X, µ), w) is path connected. It is also clear from 15.8 that the space (ERG(Γ, X, µ), w) is path connected, when Γ does not have property (T). I do not know if (ERG(Γ, X, µ), w) is path connected for any Γ. (C) We next consider connectedness properties in (A(Γ, X, µ), u), when Γ has property (T). Theorem 15.12. Let the countable group Γ have property (T). Then for any a ∈ ERG(Γ, X, µ), the path component of a in (A(Γ, X, µ), u) is exactly its conjugacy class and thus is clopen in (A(Γ, X, µ), u).
15. CONNECTEDNESS IN THE SPACE OF ACTIONS
109
Proof. By 15.6 the conjugacy class of any a ∈ A(Γ, X, µ) is contained in its uniform path component. By 14.2 the conjugacy class of a ∈ ERG(Γ, X, µ) is clopen in the uniform topology, so it is exactly equal to its uniform path component. 2 It thus follows that if an infinite Γ has property (T), (ERG(Γ, X, µ), u) and also (FRERG(Γ, X, µ), u) is far from connected. On the other hand, we have seen in 15.5 that (FRERG(Z, X, µ), u) = (ERG(Z, X, µ), u) = (ERG, u) is path connected and in fact the same argument works in the following more general situation. Theorem 15.13. Let Γ be a countable amenable group. Then the set FRERG(Γ, X, µ) is path connected in the uniform topology. Proof. As in the proof of 15.5 and using the fact that if a ∈ A(Γ, X, µ), then the equivalence relation Ea is hyperfinite (Ornstein-Weiss [OW]), it is enough to show that if a, b ∈ FRERG(Γ, X, µ) and Ea ⊆ Eb , then there is a continuous path in the space (FRERG(Γ, X, µ), u) from a to b. Now Foreman-Weiss [FW], proof of Claim 19, show that the set of conjugates of a by elements of [Eb ] contains in its uniform closure the action b. Then we can use the argument in the proof of 15.8. 2 It would be interesting to characterize the countable groups Γ for which (FRERG(Γ, X, µ), u) is path connected. Let now D(Γ, X, µ) = {a : ∃A(0 < µ(A) & A is a-invariant & a|A is ergodic)}. Thus D is the set of all actions whose ergodic decomposition has a “discrete” component, i.e., an ergodic component of positive measure. Thus ERG(Γ, X, µ) ⊆ D(Γ, X, µ). Let also C(Γ, X, µ) = A(Γ, X, µ) \ D(Γ, X, µ) be the set of all actions with “continuous” ergodic decomposition. Proposition 15.14. Let Γ be a countable group. Then C(Γ, X, µ) is dense in (A(Γ, X, µ), w). Proof. Fix b ∈ C(Γ, Y, ν). Since, by 10.4, a ∈ A(Γ, X, µ) is in the weak closure of the set of actions isomorphic to a × b, it is enough to show that the latter admits no invariant set of positive measure on which it is ergodic. Assume otherwise, and let A ⊆ X ×Y be Borel, a×b-invariant with (µ×ν)|A , π : X × Y → Y the (µ × ν)(A) > 0 and a × b|A ergodic. Let σA = (µ×ν)(A) projection, and put ρ = π∗ σA . Then ρ is a b-invariant, ergodic probability measure, so if Yρ is the ergodic component corresponding to ρ, ν(Yρ ) = 0 and ρ(Yρ ) = 1. Then (µ × ν)(X × Yρ ) = 0 and so ρ(Yρ ) = σA (π −1 (Yρ )) = µ×ν((X×Yρ )∩A) = 0, a contradiction. 2 (µ×ν)(A) Proposition 15.15. Let Γ be a countable group. Then D(Γ, X, µ) is dense in (A(Γ, X, µ), u).
110
II. THE SPACE OF ACTIONS
Proof. Let a ∈ C(Γ, X, µ). Let π : X → E be the ergodic decomposition of a, i.e., E is the standard Borel space of a-invariant, ergodic measures, π is an a-invariant Borel surjection, and for each e ∈ E, e(π −1 (e)) = 1 (and e is the unique a-invariant, ergodic measure on π −1 (e)). Let π∗ µ = ν. Since a ∈ C(Γ, X, µ), ν is non-atomic, thus we can find At ∈ MALGν , 0 ≤ t ≤ 1, with A0 = ∅, A1 = E, s ≤ t ⇒ As ⊆ At , ν(At ) = t. Let Xt = π −1 (At ). Then Xt is a-invariant, X0 = ∅, X1 = X, s ≤ t ⇒ Xs ⊆ Xt , µ(Xt ) = t. Fix now 1 > > 0 and let Y = X \ X . Fix a measure preserving, ergodic action a of Γ on (X , (µ|X )/) and let b = a|Y ∪ a . Then b ∈ D(Γ, X, µ) and clearly δΓ.u (a, b) ≤ .
2
A similar argument shows the following. Proposition 15.16. Let Γ be a countable group. Then C(Γ, X, µ) is path connected in the uniform topology and thus is contained in the path connected component of the trivial action iΓ ∈ A(Γ, X, µ) in the uniform topology. Proof. In the notation of the preceding proof, let a ∈ C(Γ, X, µ) and put at = a|Xt ∪ iΓ |Yt , 0 ≤ t ≤ 1. Then t 7→ at is a continuous path in (C(Γ, X, µ), u) which connects iΓ to a1 = a. 2 Finally we have the following result. Theorem 15.17 (Kechris). If a countable group Γ has property (T), then D(Γ, X, µ) is open (and dense) in (A(Γ, X, µ), u) and C(Γ, X, µ) is exactly the path component of iΓ in (A(Γ, X, µ), u). Proof. Let a ∈ D(Γ, X, µ) and fix an a-invariant set C with µ(C) > 0 and a|C ergodic. Next, by 14.1, there is η > 0 so that if δΓ,u (a, b) < η, then for the ϕ : A → B given in the lemma, we have µ(A) > 1 − µ(C), thus µ(A ∩ C) > 0. Since A ∩ C is a-invariant, it follows that A ∩ C = C, so C ⊆ A and since ϕ gives an isomorphism of a|A with b|B, it follows that ϕ(C) is b-invariant and b|ϕ(C) is ergodic, so b ∈ D(Γ, X, µ) (and there is b-invariant set D = ϕ(C) with µ(D) = µ(C) and b|D ergodic). This proves that D(Γ, X, µ) is open. Finally we show that C(Γ, X, µ) is the uniform path connected component of iΓ . Indeed, let t 7→ at be a uniformly continuous map from [0,1] to A(Γ, X, µ) with a0 = iΓ and a1 = a ∈ D(Γ, X, µ), towards a contradiction. Let C, η be as above and find t0 = 0 < t1 < t2 < · · · < tn = 1 such that δΓ,u (ati , ati+1 ) < η, ∀i < n. By the preceding argument, a simple backwards induction shows that for every i ≤ n, ati ∈ D(Γ, X, µ), so iΓ ∈ D(Γ, X, µ), a contradiction. 2
15. CONNECTEDNESS IN THE SPACE OF ACTIONS
111
Remark. One can also give a more direct proof for the fact that D(Γ, X, µ) is open in (A(Γ, X, µ), u), when Γ has property (T). We start with the following lemma that we will also use in 15.20. Lemma 15.18. Suppose (Q, ) is a Kazhdan pair for Γ. Let a ∈ A(Γ, X, µ), and fix δ > 0. If a Borel set A satisfies µ(γ a (A)∆A) < δ2 µ(A), ∀γ ∈ Q, then there is an a-invariant Borel set B with µ(A∆B) < 8δµ(A). Proof. Consider the Koopman representation κa associated to a. Then we have√kκa (γ)(χA ) − χA k2 = µ(γ a (A)∆A), kχA k2 = µ(A), so kκa (γ)(χA ) − 2 a ∀γ ∈ Q. Thus χA k < δkχA k, √ p there is f ∈ L (X, µ) which is κ -invariant and kχA − f k ≤ δkχA k = δµ(A). Let B = {x : |f (x) − 1| ≤ 12 }. Then B is a-invariant and 1 x ∈ A \ B ⇒ |χA (x) − f (x)| = |1 − f (x)| > , 2 R R so δµ(A) ≥ kχA − f k2 = |χA (x) − f (x)|2 dµ > A\B 41 dµ = 41 µ(A \ B). Also 1 x ∈ B \ A ⇒ |χA (x) − f (x)| = |f (x)| ≥ , 2 thus 1 δµ(A) ≥ µ(B \ A), 4 2
so 14 µ(A∆B) < 2δµ(A) or µ(A∆B) < 8δµ(A).
To complete the proof fix a ∈ D(Γ, X, µ) and let A ∈ MALGµ be a-invariant with µ(A) > 0 and a|A ergodic. Let Γ = {γ1 , γ2 , . . . }, Q = {γ1 . . . , , γN } and recall that δΓ,u (a, b) =
∞ X
2−n δu (γna , γnb ) ≥ 2−N
n=1
X
δu (γ a , γ b ).
γ∈Q
We will see that if δΓ,u (a, b) <
2−N 2 µ(A), 100
then b ∈ D(Γ, X, µ). Fix such an action b. Then µ(γ a (A)∆γ b (A)) = 2 µ(A∆γ b (A)) ≤ 2N δΓ,u (a, b) < 100 µ(A), ∀γ ∈ Q, so there is a b-invariant 8 110 set B ∈ MALGµ with µ(A∆B) < 100 µ(A), thus in particular 100 µ(A) > 90 µ(B) > 100 µ(A). If b 6∈ D(Γ, X, µ) towards a contradiction, there is a binvariant set C ⊆ B in MALGµ with µ(C) = 21 µ(B). Then µ(γ a (C)∆C) = 2 2 2 µ(γ a (C)∆γ b (C)) ≤ 2N δΓ,u (a, b) < 100 µ(A) < 90 µ(B) = 45 µ(C), ∀γ ∈ Q. 8 8 Thus there is a-invariant D ∈ MALGµ with µ(C∆D) < 45 µ(C) = 90 µ(B).
112
II. THE SPACE OF ACTIONS
Then µ(D \ A) ≤ µ(D∆C) + µ(C \ A) ≤ µ(D∆C) + µ(C \ B) + µ(A∆B) 8 8 < µ(B) + µ(A) 90 100 20 < µ(A), 100 8 37 36 µ(B) = 90 µ(B) > 100 µ(A). and µ(D) ≥ µ(C) − µ(C∆D) > 21 µ(B) − 90 1 8 53 65 Also µ(D) ≤ µ(C) + µ(D∆C) < 2 µ(B) + 90 µ(B) = 90 µ(B) < 100 µ(A). So 0 < µ(A ∩ D) < µ(A) and A ∩ D is a-invariant, which violates the ergodicity of a|A. We can in fact determine exactly the uniform path component of every a ∈ A(Γ, X, µ), using similar arguments. Theorem 15.19 (Kechris). Let Γ beSa countable group with property (T). For each a ∈ A(Γ, X, µ), let Da = {A ∈ MALGµ : µ(A) > 0, A is a-invariant and a|A is ergodic} be the “discrete” part of the ergodic decomposition of a. Then a, b are in the same uniform path connected component of A(Γ, X, µ) iff µ(Da ) = µ(Db ) and a|Da ∼ = b|Db . Proof. By 15.17 we can find a0 , uniformly path connected to a, and uniformly path connected to b, so that a|Da = a0 |Da and a0 is trivial on X \ Da and similarly for b, b0 . So we can assume that we work with a, b for which a|(X \ Da ), b|(X \ Db ) are trivial. If then a|Da ∼ = b|Db and µ(Da ) = µ(Db ), clearly a, b are conjugate, thus belong to the same uniform path connected component. Conversely, assume such a, b are uniformly path connected. Then an argument similar to the proof of 15.17 shows that there is a 1-1 correspondence between the set Da = {A ∈ MALGµ : µ(A) > 0, A is a-invariant and a|A is ergodic} and the corresponding Db , say A 7→ BA , with a|A ∼ = b|BA , from which it follows that µ(Da ) = µ(Db ) and a|Da ∼ = b|Db . 2 b0 ,
Remark. Using the last Remark of Section 14, (A), it follows that 15.19 holds for every group Γ, if the uniform topology is replaced by the δΓ,∞ -topology. (D) As we have seen in 12.2, ii), for every countable group Γ which does not have property (T), the generic element of (A(Γ, X, µ), w) is ergodic. The next result asserts that when Γ is a countable infinite group which has property (T), then the generic element of (A(Γ, X, µ), w) has continuous ergodic decomposition. Theorem 15.20 (Kechris). Let Γ be an infinite countable group with property (T). Then C(Γ, X, µ) is dense Gδ in (A(Γ, X, µ), w). Proof. By 15.14 it is enough to show that C(Γ, X, µ) is Gδ . Fix a Kazhdan pair (Q, ) for Γ. Fix also a countable dense set D ⊆ MALGµ \{∅}. Then we claim that the following are equivalent for a ∈ A(Γ, X, µ):
16. THE ACTION OF SL2 (Z) ON T2
113
(i) a ∈ C(Γ, X, µ), (ii) For every A ∈ D and for all n > 8, we have that the condition 2 ∀γ ∈ Q(µ(γ a (A)∆A) < n µ(A)) implies that there is B such that µ(B \A) < 8 1 4 1 4 a n µ(A) and ( 2 − n )µ(A) < µ(B) < ( 2 + n )µ(A) and ∀γ ∈ Q(µ(γ (B)∆B) < 2 n µ(B)), This clearly shows that C(Γ, X, µ) is Gδ . 2 (i) ⇒ (ii): Fix A ∈ D, n > 8 such that ∀γ ∈ Q(µ(γ a (A)∆A) < n µ(A)). Then by 15.18 find a-invariant A0 ∈ MALGµ with µ(A∆A0 ) < n8 µ(A). Since a ∈ C(Γ, X, µ), there is a-invariant B ∈ MALGµ , with B ⊆ A0 and µ(B) = 1 8 0 0 0 2 µ(A ). Then we have µ(B \ A) ≤ µ(A∆A ) + µ(B \ A ) < n µ(A) and 8 0 0 0 µ(A) − µ(A∆A ) ≤ µ(A ) ≤ µ(A) + µ(A∆A ), so (1 − n )µ(A) < µ(A0 ) < (1 + n8 )µ(A), therefore ( 12 − n4 )µ(A) < µ(B) < ( 21 + n4 )µ(A). Obviously 2 ∀γ ∈ Q(µ(γ a (B)∆B) < n µ(B)). (ii) ⇒ (i): Fix A0 ∈ MALGµ \ {∅} which is a-invariant. We will find a-invariant B 0 ∈ MALGµ such that 0 < µ(A0 ∩ B 0 ) < µ(A0 ). To do this, first choose a very large n (to satisfy what will be needed below). Then find A ∈ D so that µ(A∆A0 ) is very small compared with µ(A0 ) (and thus 2 µ(A)), so that, in particular, ∀γ ∈ Q(µ(γ a (A)∆A) < n µ(A)). By (ii) then there is B with µ(B \ A) < n8 µ(A), ( 12 − n4 )µ(A) < µ(B) < ( 12 + n4 )µ(A) 2 and ∀γ ∈ Q(µ(γ a (B)∆B) < n µ(B)). Finally by 15.18 find a-invariant B 0 with µ(B∆B 0 ) < n8 µ(B). It follows that µ(B 0 \ A0 ) is very small compared to µ(A0 ) and also µ(B 0 ) is very close to 21 µ(A0 ), which implies that 0 < µ(A0 ∩ B 0 ) < µ(A0 ) and completes the proof. 2 Remark. Recall that for Γ with property (T), Glasner-Weiss have shown that ESIM(Γ) is closed in SIM(Γ) (see the remarks following 12.2). The preceding result implies, in view of the result of Glasner-King (see Section 10, (B)), that the set of Borel probability measures in ESIM(Γ) which are continuous (i.e., non-atomic) is dense Gδ in the compact space of all probability measures on ESIM(Γ), which implies that ESIM(Γ) is perfect. We are using here the following result of the Choquet theory: if P (ESIM(Γ)) is the compact metrizable space of the probability measures on ESIM(Γ), with the weak∗R-topology, the map µ ∈ P (ESIM(Γ)) 7→ b(µ) ∈ SIM(Γ) R barycenter R given by f db(µ) = ( f de)dµ(e), for any f ∈ C(TΓ ), is a homeomorphism of P (ESIM(Γ)) and SIM(Γ), and for ν ∈ SIM(Γ), the unique µ with b(µ) = ν is the measure corresponding to the ergodic decomposition of ν with respect to the shift on TΓ . Remark. From 15.20 and 12.2 it follows that for any countable group Γ the set D(Γ, X, µ) \ ERG(Γ, X, µ) is meager in (A(Γ, X, µ), w). 16. The action of SL2 (Z) on T2 (A) Gaboriau and Popa [GPo] were the first to construct continuum many non-OE measure preserving, free, ergodic actions of F2 , using methods
114
II. THE SPACE OF ACTIONS
from operator algebras. Later T¨ornquist [To1] used some ideas in the work of Popa [Po2] and Gaboriau-Popa together with a genericity argument to find a more elementary proof and established a descriptive strengthening of their result, which we will discuss in Section 17 (in a somewhat stronger form). A basic idea, going back to Popa, is to employ the standard action of SL2 (Z) (and its free subgroups of finite index) on T2 . We will first recast below some important properties of this action in a somewhat different form that brings out more clearly its essential features. We will then derive in the next section the corollaries concerning orbit equivalence. (B) Consider the action of SL2 (Z) on (X, µ) = (T2 , µ), where µ is the usual product measure on T2 = R2 /Z2 , given by the following matrix multiplication: −1 t z1 . A · (z1 , , z2 ) = (A ) z2 Fix also a copy of F2 in SL2 (Z) which has finite index in SL2 (Z), and consider the induced action a0 of F2 on (X, µ). This is free, measure preserving, ergodic (in fact weak mixing and even tempered (but not mixing), see Kechris [Kec4] 5, (C)). Denote by E a 0 = F0 the equivalence relation induced by this action. We will see that F0 has the property that all its measure preserving extensions E ⊇ F0 exhibit very interesting properties. To formulate this, define the following left-invariant metric on Aut(X, µ): X d0 (S, T ) = δw (S, T ) + 2−n δu (Sγn S −1 , T γn T −1 ) n
= δw (S, T ) + δF2 ,u (Sa0 S −1 , T a0 T −1 ), where F2 = {γn } and γ ∈ F2 is identified here with γ a0 . Note that if E, F ⊇ F0 are measure preserving countable Borel equivalence relations, and T : E ∼ = F is an isomorphism, then T γT −1 ∈ [F ], so ifSN [E, F ] denotes the set of all T : E ∼ = F , then for each such F ⊇ F0 , d0 | {N [E, F ] : E ∼ = F, E ⊇ F0 } is separable. We can now recast the crucial properties of this action in the following lemma. Theorem 16.1 (Popa). For some δ > 0 and any E, F ⊇ F0 , T ∈ N [E, F ], we have: d0 (T, 1) < δ ⇒ δu (T, 1) < 1 ⇒ E = F. This has the following immediate corollary, noticing that if E = F , N [E, F ] = N [E] and d0 |N [E] is continuous in the topology of N [E]. Corollary 16.2. For any E ⊇ F0 , [E] is clopen in N [E] and thus Out(E) is countable.
16. THE ACTION OF SL2 (Z) ON T2
115
Proof. Taking E = F in 16.1, we see that there is > 0 such that if T ∈ N [E] and dN [E] (T, 1) < , then d0 (T, 1) < δ, so δu (T, 1) < 1, i.e., T (x) = x on a set of positive measure, thus (since clearly E is ergodic) T ∈ [E]. So [E] is clopen in N [E]. 2 Corollary 16.3. On the set of measure preserving countable Borel equivalence relations E ⊇ F0 , isomorphism, E ∼ = F , is a countable equivalence relation (i.e., every equivalence class is countable). Proof. Fix E ⊇ F0 and assume, towards a contradiction, that {Ei }i∈I is an uncountable family of distinct equivalence relations S with Ei ⊇ F0 and E ∼ = E. Since d0 | F ∼ = Ei . Fix isomorphisms Ti : Ei ∼ =E,F ⊇F0 N [F, E] is separable, it follows that d0 |{Ti : i ∈ I} is separable, so there are i 6= j in I with d0 (Ti , Tj ) < δ (where δ is as in 16.1). Thus d0 (Ti−1 Tj , 1) < δ and, since Ti−1 Tj ∈ N [Ej , Ei ], this implies, by 16.1, that Ej = Ei , a contradiction. 2 Proof of 16.1. First we prove the implication δu (T, 1) < 1 ⇒ E = F. If δu (T, 1) < 1, then if A = {x : T (x) = x}, µ(A) > 0. Thus the F2 saturation of A is equal to X. Suppose that xEy. Then for some γ, δ ∈ F2 , γ · x, δ · y ∈ A, and (γ · x)E(δ · y), as F0 ⊆ E. Then T (γ · x) = (γ · x)F (δ · y) = T (δ · y), so xF y, as F0 ⊆ F . Thus E ⊆ F and similarly F ⊆ E. We will now prove that there is δ such that d0 (T, 1) < δ ⇒ δu (T, 1) < 1. Below we will need some facts concerning the so-called relative property (T) for pairs of groups. These can be all found in [BdlHV], 1.4 and Jolissaint [Jo3]. The crucial point is that if we consider the usual action of SL2 (Z) on Z2 by matrix multiplication and the corresponding semi-direct product G1 = SL2 (Z) n Z2 , then (G1 , Z2 ) has the relative property (T), i.e., there is finite Q ⊆ G1 and > 0 such that in any unitary representation π of G1 on a Hilbert space H, if there is a (Q, )-invariant vector v, i.e., kπ(γ)(v) − vk < kvk, ∀γ ∈ Q, then there is Z2 -invariant non-0 vector. If G = F2 n Z2 , then (G, Z2 ) also has the relative property (T) and by standard facts this also implies the following, which is all we will use in the sequel: (*) There is finite Q ⊆ F2 n Z2 and > 0 such that for any unitary representation π of F2 n Z2 on a Hilbert space H, if ξ is a (Q, )-invariant unit vector, then there is a Z2 -invariant vector η with kξ − ηk < 1. c2 of T2 , m = We will identify below Z2 with the group of characters T 2 (m1 , m2 ) ∈ Z being identified with the character χm (z1 , z2 ) = z1m1 z2m2
116
II. THE SPACE OF ACTIONS
(where we view here T as the multiplicative group of the unit circle). Then the action of SL2 (Z) (and thus of F2 ≤ SL2 (Z)) on T2 gives rise to an c2 given by g · χ(z) = χ(g −1 · z) and by the above identification action on T this is exactly the matrix multiplication action of SL2 (Z) on Z2 . Thus the semidirect product F2 n Z2 can be viewed (see Appendix I, (B)) as the set c2 and multiplication defined by of pairs (γ, χ), with γ ∈ F2 , χ ∈ T (γ1 , χ1 )(γ2 , χ2 ) = (γ1 γ2 , (γ2−1 · χ1 )χ2 ). Fix now E, F ⊇ F0 , T ∈ N [E, F ] and let (Q, ) be as in (*) above. We will use T to define a unitary representation π of F2 n Z2 on H = L2 (F, M ) (where M is defined as in Section 6, (B)). Letting π(g)(f ) = g · f , this is defined by (γ, χ) · f (z, w) = χ(γ −1 (z))χ(γ −1 T −1 (w))f (γ −1 (z), T γ −1 T −1 (w)), for f ∈ L2 (F, M ), where again γ ∈ F2 is identified with γ a0 . (Note that this is well-defined as zF w ⇔ γ −1 (z)F T γ −1 T −1 (w).) It is straightforward to check that this is an action. Claim. We can find δ > 0 such that if d0 (T, 1) < δ, then ξ = ∆ = (the characteristic function of the diagonal) is (Q, )-invariant. Granting this, there is a Z2 -invariant vector η in L2 (F, M ) with kξ−ηk < 1 and thus η(z, z) 6= 0 for a positive set of z, say A. Since (1, χ) · η = η for c2 , we have all χ ∈ T η(z, z) = χ(z)χ(T −1 (z))η(z, z), thus c2 , χ(T −1 (z)) = χ(z), ∀z ∈ A, ∀χ ∈ T so, as the characters separate points, T −1 (z) = z and so T (z) = z on A, i.e., δu (T, 1) < 1. Proof of the claim. We have (γ, χ) · ∆(z, w) = χ(γ −1 (z))χ(γ −1 T −1 (w))∆(γ −1 (z), T γ −1 T −1 (w)), thus k(γ, χ) · ∆ − ∆k22 is equal to Z 2 χ(γ −1 (z))χ(γ −1 T −1 (w))∆(γ −1 (z), T γ −1 T −1 (w)) − ∆(z, w) dM F
and thus to Z 2 (1) χ(γ −1 (z))χ(γ −1 T −1 (z))∆(γ −1 (z), T γ −1 T −1 (z)) − 1 dµ(z) 2 T Z (2) + ∆(γ −1 (z), T γ −1 T −1 (w))dM. z6=w,(z,w)∈F
Now (2) is equal to δu (γ −1 , T γ −1 T −1 ) = δu (γ, T γT −1 ),
17. NON-ORBIT EQUIVALENT ACTIONS OF FREE GROUPS
117 2
so it is clear that if δ is small enough and d0 (T, 1) < δ, then (2)< 2 for all (γ, χ) ∈ Q. Also note that for (1) (considering cases whether γ −1 (z) = T γ −1 T −1 (z) or not), we see that (1) ≤ δu (γ −1 , T γ −1 T −1 ) + k(γ · χ) − UT (γ · χ)k2 , where γ · χ(z) = γ(χ−1 (z)) and UT is the unitary operator on L2 (T2 , µ) corresponding to T (i.e., UT (f )(z) = f (T −1 (z)). Now d0 (T, 1) ≥ δw (T, 1) and the topology induced by δw is the weak topology, i.e., the topology induced by identifying T and UT and using on U (L2 (T2 , µ)) the strong operator topology. So if δ is small enough, we have again d0 (T, 1) < δ ⇒ 2 (1) < 2 , ∀(γ, χ) ∈ Q, and thus finally d0 (T, 1) < δ ⇒ k(γ, χ) · ∆ − ∆k2 < , for all (γ, χ) ∈ Q and the proof is complete.
2
17. Non-orbit equivalent actions of free groups (A) We will now prove the following non-classification result. The first assertion is in T¨ornquist [To1] and the second can be obtained by combining his method with 5.5. Theorem 17.1 (T¨ ornquist). (a) The equivalence relation E0 can be Borel reduced to the orbit equivalence relation of measure preserving, free, ergodic actions of the free group F2 . (b) This orbit equivalence relation cannot be classified by countable structures. Proof. We fix, by the work in Section 16, a measure preserving, free, ergodic action a0 of F2 on (X, µ) such that if F0 = Ea0 is the induced equivalence relation, then isomorphism on the set of all E ⊇ F0 is a countable equivalence relation. Let F2 = hα, βi, F3 = hα, β, γi be free generators. Lemma 17.2. Fix a0 ∈ FR(F2 , X, µ). Then for a comeager set of S ∈ Aut(X, µ) the action aS of F3 determined by αaS = αa0 , β aS = β a0 , γ aS = S is free (and of course ergodic). We will grant this and complete the proof of 17.1. We will first prove the version of 17.1 for F3 instead of F2 . Consider the equivalence relation ∼ on Aut(X, µ) defined by S ∼ T ⇔ aS OEaT . To prove (b), note that if orbit equivalence of measure preserving, free, ergodic actions of F3 can be classified by countable structures, so can ∼ and we will derive a contradiction from this. From 5.5, we know that the conjugacy action of ([F0 ], u) on the space (Aut(X, µ), w) is generically turbulent. Let ∼C be the corresponding equivalence relation. Then clearly S ∼C T ⇒ EaS = EaT ⇒ aS OEaT ⇔ S ∼ T.
118
II. THE SPACE OF ACTIONS
So if f : Aut(X, µ) → XL , f Borel and L some countable language, is such that S ∼ T ⇔ f (S) ∼ = f (T ), then S ∼C T ⇒ f (S) ∼ = f (T ), so, by turbulence, there is a comeager set A and M0 ∈ XL with ∼ M0 , ∀S ∈ A. f (S) = Now we claim that every ∼-class is meager, which implies that there are S, T ∈ A with S 6∼ T and thus f (S) ∼ 6 f (T ), a contradiction. To prove this = let S ≈ T ⇔ Ea S = Ea T . Then by the above property of F0 (since EaS , EaT ⊇ F0 ), we have that each ∼-class contains only countably many ≈-classes, so it is enough to check that each ≈-class is meager. But this is obvious, as [S]≈ ⊆ [EaS ], which is a meager set in (Aut(X, µ), w). To prove (a) it is enough to show that E0 is Borel reducible to ∼. Now ∼ is a meager equivalence relation (being analytic with meager classes) and contains the equivalence relation induced by the conjugacy action of ([F0 ], u) on (Aut(X, µ), w), which clearly has a dense orbit. So E0 is Borel reducible to ∼ by the argument in Becker-Kechris [BK], 3.4.5 (see also Hjorth-Kechris [HK1], 3.2). To prove the result for F2 , it is enough to show that OE for measure preserving, free, ergodic actions of F3 can be Borel reduced to OE for measure preserving, free, ergodic actions of F2 . To see this, fix (X, µ) and let Y = X × {0, 1}, ν = µ × p, where p({0}) = p({1}) = 1/2. Let τ ∈ Aut(Y, ν) be defined by τ (x, i) = (x, 1 − i). Each S ∈ Aut(X, µ) induces S˜ ∈ Aut(Y, ν) ˜ 0) = (S(x), 0), S(x, ˜ 1) = (x, 1). Given a ∈ FRERG(F3 , X, µ), by S(x, ˜ let Ea be the equivalence relation on Y induced by α˜a , β˜a , γ˜a , τ , where F3 = hα, β, γi. Then it is not hard to see that ˜a ∼ ˜b . Ea ∼ = Eb ⇔ E =E ˜a is ergodic. Finally, it is also not hard to see that E ˜a is induced Clearly E by a free, measure preserving action a ˜ of F2 = hα, βi, namely ( τ (x, i), if i = 0, αa˜ (x, i) = fa τ (x, i), if i = 1, α ( fa (x, i), if i = 0, τγ β a˜ (x, i) = fa β τ (x, i), if i = 1. So it only remains to give the proof of 17.2. Proof of 17.2. Let T1 = αa0 , T2 = β a0 . For each > 0, and non-trivial reduced word v ∈ F3 , the set {S : δu (v T1 ,T2 ,S , 1) > 1 − }
17. NON-ORBIT EQUIVALENT ACTIONS OF FREE GROUPS
119
is open in (Aut(X, µ), w), so, by the Baire Category Theorem, it is enough to show that it is dense. By induction, we can assume that the intersection of the following sets, where v¯ varies over all words of length less than that of v, {S : δu (¯ v T1 ,T2 ,S , 1) = 1}, is dense in (Aut(X, µ), w). Put for convenience for each word ω, ω S = ω T1 ,T2 ,S . Finally, we can assume that γ occurs in v. Let v0 , v1 , v2 , . . . , vn = v, where n = length(v), be the terminal subwords of v with length(vi ) = i, and let i0 < n be largest such that vi0 +1 = γ ±1 vi0 . Assume, without loss of generality that vi0 +1 = γvi0 , so that viS0 +1 = SviS0 . We will use the following two simple lemmas. Lemma 17.3. If P1 , . . . , Pk ∈ Aut(X, µ) and µ({x : P1 (x), . . . , Pk (x) are distinct} ) > a, then there are pairwise disjoint Borel sets A1 , . . . , Am with µ(A1 ∪ . . . ∪Am ) > a and P1 (Ai ), . . . , Pk (Ai ) pairwise disjoint, for each 1 ≤ i ≤ m. Proof. As in the proof of 3.10, there is a countable collection of Borel sets {Un } such that for each x with P1 (x), . . . , Pk (x) distinct, there is Un(x) with x ∈ Un(x) and P1 (Un(x) ), . . . , Pk (Un(x) ) pairwise disjoint. Thus, if S A = {x : P1 (x), . . . , Pk (x) are distinct}, then A = i∈N Ai , where the sets P1 (Ai ), . . . , Pk (Ai ) are pairwise disjoint, from which the conclusion follows. 2 Lemma 17.4. For any T ∈ Aut(X, µ) and any Borel sets A, A1 , . . . , Am , there is S ∈ Aut(X, µ), such that S(Ai ) = T (Ai ), S −1 (Ai ) = T −1 (Ai ) ∀i ≤ m, and {x : S(x) 6= T (x)} = A. Proof. Considering the partition generated by A, A1 , . . . , Am , T −1 (A1 ), . . . , T −1 (Am ) and the fact that for each Borel set B there is an involution PB ∈ Aut(X, µ) with supp(PB ) = B, it follows that there is an involution P ∈ Aut(X, µ) with supp(P ) = A and P (Ai ) = Ai , P (T −1 (Ai )) = T −1 (Ai ), ∀i ≤ m. Let then S = T P . 2 Continuing the proof, it is enough to show that for each S0 such that δu (¯ v S0 , 1) = 1 for every word v¯ of length < n, and each weak nbhd N0 of S0 , N0 ∩ {S : δu (v S , 1) > 1 − } = 6 ∅. SBy 17.3, we can find pairwise disjoint Borel sets A1 , . . . , Am ⊆ X with µ( m i=1 Ai ) > 1 − and viS0 (Ai ) ∩ vjS0 (Ai ) = ∅, i 6= j < n. If δu (v S0 , 1) > 1 − , there is nothing to prove. Else (renumbering the Ai ’s if necessary) on a set of positive measure in A1 we have v S0 (x) = x.
120
II. THE SPACE OF ACTIONS
Put A01 = {x ∈ A1 : v S0 (x) = x}. Then, by 17.4, find S1 ∈ N0 such that {x : S1 (x) 6= S0 (x)} = viS00 (A01 ). It follows that v S1 (x) 6= x for x ∈ A1 . (Note that S1 = S0 , S1−1 = S0−1 on the sets A1 , v1S0 (A1 ), . . . , viS00−1 (A1 ).) Then, by Lemma 17.3 again, we can find pairwise disjoint B1 , . . . , Bk ⊆ A1 such that S µ( kj=1 Bj ∪ A2 ∪ · · · ∪ Am ) > 1 − and v S1 (Bj ) ∩ Bj = ∅, ∀j ≤ k. If δu (v S1 , 1) > 1−, then again there is nothing to prove, so (renumbering again the Ai ’s, i ≥ 2, if necessary), there is a set of positive measure in A2 for which v S1 (x) = x. Put A02 = {x ∈ A2 : v S1 (x) = x}. As before we can find S2 ∈ N0 such that {x : S2 (x) 6= S1 (x)} = viS01 (A02 ) and S2±1 (viS1 (Bj )) = S1±1 (viS1 (Bj )), ∀i < n, j ≤ k, so that viS2 (Bj ) = viS1 (Bj ), ∀i ≤ n, j ≤ k, thus v S2 (Bj ) ∩ Bj = ∅, ∀j ≤ k. Also, as before, v S2 (x) 6= x for x ∈ A2 , so we can S S find pairwise disjoint C1 , . . . , C` ⊆ A2 with µ( kj=1 Bj ∪ `t=1 Ct ∪ A3 ∪ · · · ∪ Am ) > 1− and v S2 (Bj )∩Bj = v S2 (Ct )∩Ct = ∅, ∀j ≤ k, ∀t ≤ `. Proceeding ¯ this way, finitely many times, we can find S¯ ∈ N0 with δu (v S , 1) > 1 − and the proof is complete. 2 (B) In connection with 17.1, we note that it would also follow from 13.5 if one knew that orbit equivalence classes in ERG(F2 , X, µ) were meager in (ERG(F2 , X, µ), w). So we can raise the following general question (analogous to that of 13.6 for orbit equivalence). Problem 17.5. For which infinite countable groups Γ are the orbit equivalence classes in ERG(Γ, X, µ) meager in the weak topology of ERG(Γ, X, µ)? Dye’s theorem 3.13 and the Ornstein-Weiss [OW] theorem that every measure preserving action of a countable amenable group gives rise to a hyperfinite equivalence relation, imply that there is only one orbit equivalence class in ERG(Γ, X, µ), when Γ is infinite amenable, so 17.5 has clearly a negative answer for Γ amenable. Finally, in view of 14.4 and 17.1, we have the following problem. Problem 17.6. Let Γ be a non-amenable group. Is it true that orbit equivalence of free, measure preserving, ergodic actions of Γ cannot be classified by countable structures? Addendum. Recently Adrian Ioana [I1] has shown that any group Γ with F2 ≤ Γ has continuum many non-orbit equivalent, free, measure preserving, ergodic actions. In fact using his main lemma one can derive a positive answer to 17.6 for such groups, and also show that E0 can be Borel reduced to orbit equivalence. We outline the argument below. Consider the action a0 ∈ A(F2 , X, µ), where X = T2 , µ = the usual product measure on T2 , defined in Section 16, (B). This action is free and weak mixing. Let also Γ be a countable group with F2 ≤ Γ. The following is proved in Ioana [I1], 1.3:
17. NON-ORBIT EQUIVALENT ACTIONS OF FREE GROUPS
121
Consider the set A ⊆ A(Γ, Y, ν) ((Y, ν) some fixed space) of all a ∈ FR(Γ, Y, ν) satisfying the following conditions: (i) a|F2 ∈ ERG(F2 , Y, ν), (ii) There is a Borel map ϕ : Y → X witnessing that a0 is a factor of a|F2 such that for each γ ∈ Γ, γ 6= 1, {y ∈ Y : π(y) = π(γ a (y))} is ν-null. Define the equivalence relation R on A by: aRb ⇔ a|F2 ∼ = b|F2 . Then on the set A, OE has countable index over OE∩R (i.e., every OE class in A contains only countably many OE∩R classes). One can also see, using, for example, Ioana [I1], 2.2 (i) that CIndΓF2 (a0 )|F2 is weak mixing. Let now π ∈ Irr(F2 , H), π ≺ λF2 (see Appendix H). Let π ˜ = IndΓF2 (π) be the induced representation. Since IndΓF2 (λF2 ) ∼ = λΓ and Γ Γ π ≺ λF2 ⇒ IndF2 (π) ≺ IndF2 (λF2 ), it follows that π ˜ ≺ λΓ (see Bekka-de la Harpe-Valette [BdlHV]). Moreover π ˜ |∆ ≺ λΓ |∆ ∼ λ∆ , for any ∆ ≤ Γ, so π ˜ |F2 ≺ λF2 and thus π ˜ |F2 is weak mixing (since λF2 does not weakly contain a finite-dimensional representation; see Dixmier [Di], 18.9.5, 18.3.6). Let now aπ˜ be as in Theorem E.1, so that aπ˜ |F2 ∈ WMIX(F2 , X, µ) (see the paragraph following E.1). Put b(π) = CIndΓF2 (a0 ) × aπ˜ . Then, recalling Appendix H, (C), b : S(λF2 ) → FR(Γ, Y, ν) (for some appropriate (Y, ν)) is a Borel map (where A(Γ, Y, ν) has the weak topology) and b(π) ∈ A. Thus b−1 (OE) has countable index over b−1 (OE ∩ R) and we next verify that b−1 (OE ∩ R) has countable index over ∼ =. Indeed, if π, ρ ∈ S(λF2 ) aρ˜ b(ρ) and b(π)Rb(ρ), then ρ ≤ ρ˜|F2 ≤ κ0 |F2 ≤ κ0 |F2 (as aρ˜|F2 v b(ρ)|F2 ) ∼ = b(π) κ0 |F2 . Thus fixing π, all such ρ are irreducible subrepresentations of b(π) κ0 |F2 , so there are only countably many, up to ∼ =. It follows that b−1 (OE) has countable index over ∼ = in the space S(λF2 ). Now the conjugacy action of U (H) on S(λF2 ) is turbulent by Appendix H, (C) (see the second to the last remark). As in the proof of 17.1, this shows that E0 can be Borel reduced to OE in FRERG(Γ, X, µ) and also that the latter cannot be classified by countable structures. Addendum. Inessa Epstein [E] has now shown that any non-amenable group Γ has continuum many non-orbit equivalent, free, measure preserving, ergodic actions. By combining her work with that of Ioana-Kechris-Tsankov [IKT] one can now obtain a complete solution to Problem 17.6. We have in fact the following stronger result. Theorem 17.7 (Epstein-Ioana-Kechris-Tsankov; see [IKT]). Let Γ be a non-amenable group. Then E0 can be Borel reduced to the orbit equivalence relation of measure preserving, free, mixing actions of Γ and this orbit equivalence relation cannot be classified by countable structures.
122
II. THE SPACE OF ACTIONS
Thus we have a very strong dichotomy: If an infinite group Γ is amenable, there is exactly one orbit equivalence class of measure preserving, free, ergodic actions of Γ. If Γ is not amenable, orbit equivalence of such actions (even if they are mixing) is unclassifiable in a very strong sense. 18. Classifying group actions: A survey In several places before we have considered various notions of equivalence of actions of a given group Γ and discussed a number of results concerning the complexity of classification under each notion of equivalence. Our aim in this section is to summarize what seems to be known in this context and collect problems that are left open. Fix a countable infinite group Γ. In the space A(Γ, X, µ) we have considered the following equivalence relations: (i) a ∼ = b (isomorphism) (ii) aOEb (orbit equivalence) (iii) a ∼ =w b (weak isomorphism) a (iv) κ0 ∼ = κb0 (unitary equivalence) (v) a ∼ b (weak equivalence) These are discussed in Section 10. It is clear that (i) ⇒ (ii), (i) ⇒ (iii) ⇒ (iv), (iii) ⇒ (v). To avoid trivialities, we will consider these equivalence relations on the space FRERG(Γ, X, µ) of free, ergodic actions of Γ. We will start with Γ = Z, then consider abelian Γ, amenable Γ and general (infinite) Γ, as the information available is decreasing as we pass from one case to the next. For each group Γ, we will consider relations between (i)–(v), the Borel complexity of each one of them (in the weak topology) and finally their complexity in the hierarchy of equivalence relations. Case I. Γ = Z. (Ia) Comparisons. By Dye’s Theorem, any two a, b ∈ FRERG(Z, X, µ) = ERG(Z, X, µ) = ERG are orbit equivalent, so clearly (ii) 6⇒ (i). It is known that (iii) 6⇒ (i); see Glasner [Gl2], 7.16. If a ∼ =w b, then a, b have the same entropy, so, considering Bernoulli shifts of different entropy, we see that (iv) 6⇒ (iii). Finally, we have again that a ∼ b for any a, b ∈ ERG(Z, X, µ) (see 2.4), so (v) 6⇒ (iii). (Ib) Borel complexity. Foreman-Rudolph-Weiss [FRW] have shown that ∼ = on ERG is not Borel (it is clearly analytic). (ii), (v) are of course trivial on ERG and ∼ =w is analytic. Open problem: Is ∼ =w on ERG Borel? Finally, by spectral theory, (iv) is Borel. (Ic) Equivalence relation complexity. In terms of Borel reducibility, a (strict) lower bound for ∼ = is measure equivalence (see 5.7) and an obvious upper bound is the universal equivalence relation induced by a Borel action of the group Aut(X, µ) (or equivalently induced by a Borel action of U (H)); see Becker-Kechris [BK] for a discussion of these universal equivalence relations. Open problem. Is ∼ = on ERG Borel bireducible to the universal equivalence relation induced by a Borel action of Aut(X, µ)? We have seen
18. CLASSIFYING GROUP ACTIONS: A SURVEY
123
in 5.7 that (iv) is Borel bireducible to measure equivalence and of course (ii), (v) are trivial. Finally ∼ =w is not classifiable by countable structures; see 13.9. Open problem. Does E1 Borel reduce to ∼ =w on ERG? (See 13.10.) Remark. One can also ask about the complexity of isomorphism within each unitary equivalence class of ergodic transformations. Not much seems to be known about this problem; see Section 5, (C). Case II. Γ abelian. (IIa) Comparisons. Same as with Γ = Z, except that I do not know if (iii) ⇒ 6 (i) has been verified for all abelian Γ. (IIb) Borel complexity. Same as with Γ = Z, except that we have an additional open problem. Open problem. Is ∼ = on FRERG(Γ, X, µ) nonBorel? (IIc) Equivalence relation complexity. Same as with Γ = Z. Case III. Γ amenable. (IIIa) Comparisons. Same as with Γ abelian. (IIIb) Borel complexity. Same as with Γ abelian, except that we have an additional open problem. Open problem. Is unitary equivalence on FRERG(Γ, X, µ) Borel? (IIIc) Equivalence relation complexity. Concerning isomorphism in the space FRERG(Γ, X, µ), we have seen in 13.7 that it is not classifiable by countable structures. Since this is true even allowing C-measurable functions, it follows from Hjorth-Kechris [HK1], 2.2 that E0 can be Borel reduced to ∼ = on FRERG(Γ, X, µ). (Recall that a function between standard Borel spaces is C-measurable if the preimage of each Borel set is in the smallest σ-algebra containing the Borel sets and closed under the Souslin operation A.) Not much else seems to be known about the equivalence relation complexity of ∼ =. Again OE is trivial and so is (v). Nothing seems to be known about weak isomorphism and unitary equivalence. Case IV. General Γ. (IVa) Comparisons. As far as I know, it is open whether for every infinite countable group Γ the reversals of the implications (i) ⇒ (ii), (i) ⇒ (iii) ⇒ (iv), (iii) ⇒ (v) fail. Addendum. Kida [Ki] provides examples of groups Γ for which (i) ⇔ (ii). Also a recent result of Bowen [Bo] shows that for any Γ ⊇ F2 , (iii) ⇒ (i) fails. (IVb) Borel complexity. It is again open whether the relations (i)–(iv) are non-Borel for every countable group Γ. Clearly weak equivalence is Gδ . Addendum. T¨ornquist [To3] has recently shown that isomorphism and orbit equivalence are not Borel for groups Γ with relative property (T) over an infinite subgroup.
124
II. THE SPACE OF ACTIONS
(IVc) Equivalence relation complexity. As in (IIIc), ∼ = on the space FRERG(Γ, X, µ) cannot be classified by countable structures and E0 can be Borel reduced to ∼ = on FRERG(Γ, X, µ). Concerning orbit equivalence, it is now known (see 17.7) that for non-amenable Γ, E0 can be Borel reduced to OE on FRERG(Γ, X, µ) and the latter is not classifiable by countable structures. Nothing much is known about weak isomorphism. Concerning unitary equivalence, we have seen in 13.8 that it cannot be classified by countable structures and again it follows that E0 can be Borel reduced to it. Finally, weak equivalence, is Gδ and thus smooth. It does not seem to be known however whether for every non-amenable Γ, there is more than one weak equivalence class in FRERG(Γ, X, µ) (see the fourth remark of Section 13, (A)).
CHAPTER III
Cocycles and cohomology 19. Group-valued random variables Let G be a Polish group and dG = d a compatible metric bounded by 1. Let (X, µ) be a standard measure space. We denote by L(X, µ, G) the space of all Borel (equivalently µ-measurable) maps f : X → G, where we identify two such maps if they agree µ-a.e. Thus L(X, µ, G) is the space of G-valued random variables. We endow L(X, µ, G) with pointwise multiplication f g(x) = f (x)g(x), under which it is clearly a group. We next define a metric dX,µ,G on L(X, µ, G) by Z dX,µ,G (f, g) =
d(f (x), g(x))dµ(x).
Convention. When G is a countable (discrete) group, we will always assume that it is equipped with the metric dG = d, where d(x, y) = 1, if x 6= y. It follows that when G is countable, dX,µ,G (f, g) = µ({x : f (x) 6= g(x)}). The following proposition is obvious. Proposition 19.1. If d is left-invariant (resp., right-invariant, (two-sided) invariant), so is dX,µ,G . We denote by τ (dX,µ,G ) the topology on L(X, µ, G) induced by dX,µ,G . Proposition 19.2. τ (dX,µ,G ) is a group topology. Proof. To simplify the notation, write below d˜ = dX,µ,G . ˜ n , f ) → 0, d(g ˜ n , g) → 0 in order to show that d(f ˜ −1 , f −1 ) → Assume d(f n ˜ 0, d(fn gn , f g) → 0. Put rn (x) = d(fn (x), f (x)), sn (x) = d(fn (x)−1 , f (x)−1 ). Then rn ≥ 0, rn ∈ L1 (X, µ, R), rn → 0 in L1 (X, µ, R), so ∃{ni } such that rni (x) → 0, µ-a.e. Thus sni (x) → 0 µ-a.e., so by Lebesgue Dominated ˜ −1 , f −1 ) → 0. Similarly, for any sequence {mi } there is a Convergence d(f ni ˜ −1 , f −1 ) → 0, so d(f ˜ −1 , f −1 ) → 0. The argument subsequence {ni } with d(f ni n for multiplication is essentially the same. 2 125
126
III. COCYCLES AND COHOMOLOGY
Proposition 19.3. The topology τ (dX,µ,G ) depends only on the topology of G and the measure class of µ, i.e., if d1 , d2 ≤ 1 are compatible metrics on G and µ1 ∼ µ2 , then τ ((d1 )X,µ1 ,G ) = τ ((d2 )X,µ2 ,G ). Proof. First we show that for any µ, τ ((d1 )X,µ,G ) = τ ((d2 )X,µ,G ). Fix > 0 and find 0 < δ < such that Bd1 (1, δ/2) ⊆ Bd2 (1, /2). Then we claim that (in the notation of the proof of 19.2) Bd˜1 (1, (δ/2)2 ) ⊆ Bd˜2 (1, ). Indeed, R if d˜1 (1, f ) = d1 (1, f (x))dµ(x) < (δ/2)2 , then, by Chebychev’s inequality, µ({x : d1 (1, f (x)) ≥ δ/2}) < δ/2, so d˜2 (1, f ) ≤ δ/2 + δ/2 < . Next we show that for any d, τ (dX,µ1 ,G ) = τ (dX,µ2 ,G ). Fix again > 0 and find 0 < δ < such that for any Borel set B, µ1 (B) < δ/2 ⇒ µ2 (B) < /2. Then, as before, Bd˜µ (1, (δ/2)2 ) ⊆ Bd˜µ (1, ) (where d˜µi = dX,µi ,G ). 2 1
2
From now on we will simply write τX,µ,G for this topology. We will next verify that it coincides with the topology of convergence in measure. Recall that for fn , f ∈ L(X, µ, G), fn → f in measure if ∀ > 0(µ({x : d(fn (x), f (x)) ≥ }) → 0). Proposition 19.4. The following are equivalent for fn , f ∈ L(X, µ, G): (i) fn → f (in τX,µ,G ), (ii) fn → f in measure, (iii) For every sequence {mi }, there is a subsequence {ni } such that fni (x) → f (x), µ-a.e. (x). Proof. (i) ⇒ (ii): By Chebychev’s inequality, · µ({x : d(f (x), g(x)) ≥ }) ≤ dX,µ,G (f, g). (ii) ⇒ (i): Let rn (x) = d(fn (x), f (x)). If fn → f in measure, then rn : X → R converges in measure to 0, so there is {ni } such that rni → 0 a.e., thus by Lebesgue Dominated Convergence, dX,µ,G (fni , f ) → 0. Similarly for any sequence {mi }, there is a subsequence {ni } such that dX,µ,G (fni , f ) → 0, so fn → f . This argument also shows that (ii) ⇒ (iii). Finally, (iii) ⇒ (i) follows immediately by Lebesgue Dominated Convergence. 2 Since every two standard measure spaces (X1 , µ1 ), (X2 , µ2 ) are isomorphic, it follows that (L(X, µ, G), τX,µ,G ) (resp., (L(X, µ, G), dX,µ,G )) is determined uniquely up to homeomorphism (resp., isometry). Thus, if we do not need to indicate (X, µ) explicitly, we will sometimes write ˜ = L(X, µ, G), G τ˜ = τX,µ,G , d˜ = dX,µ,G . Next we note that the map g ∈ G 7→ (x 7→ g) is an isometric embedding ˜ so we can view G as a from G to the subgroup of constant functions in G, ˜ ˜ ˜ (topological) subgroup of G, G ≤ G, and d as an extension of d.
19. GROUP-VALUED RANDOM VARIABLES
127
n
Consider now for n ≥ 1 the product group G2 with compatible metric n
dn ({gi }, {hi }) = 2
−n
2 X
d(gi , hi ).
i=1 n
We can embed isometrically G2 into G2
n+1
via
(g1 , . . . , g2n ) → (g1 , g1 , g2 , g2 , . . . , g2n , g2n ), S n so we can write G1 ≤ G2 ≤ G4 ≤ . . . and let G(∞) S = n G(2 ) , which is a separable metrizable group with compatible metric n dn . Since (X, µ) is a standard measure space, we can find a sequence of finite Borel partitions of X, P1 ≥ P2 ≥ . . . , with Pn+1 refining Pn , Pn = S (n) (n) (n) {A1 , . . . , A2n }, µ(Ai ) = 2−n , and the Boolean algebra generated by n Pn ˜ n the subgroup of G ˜ consisting of all functions dense in MALGµ . Denote by G which are constant on each piece of the partition Pn . It is then clear that the n ˜G ˜ n are isometrically (with respect to d2n , d| ˜ n ) isomorphic, groups G2 and G n 2 (∞) ˜ ˜ n and thus G ˜ (∞) = S G so we can identify G and G with G n n. ˜ ∞ is dense in G ˜ and thus G ˜ is separable. Proposition 19.5. G ˜ Proof. Using the separability of G, it is easy to check that the f ∈ G ˜ from which it immediately follows that with countable range are dense in G, ˜ ˜ Given such an f with range the f ∈ G with finite range are dense in G. −1 {g1 , . . . , gk }, let Ai = f ({gi }), and for each > 0 find n large enough so that there is a partition B1 , . . . , Bk consisting of elements of the Boolean ˜ is defined algebra generated by Pn , with µ(Ai ∆Bi ) < /k, ∀i ≤ k. If g ∈ G ˜ ˜ by g(x) = gi , for x ∈ Bi , then g ∈ Gn and d(f, g) < . 2 Finally we have: Proposition 19.6. If the metric d on G is complete, so is the metric d˜ on ˜ Thus G ˜ is a Polish group. G. R ˜ Proof. Let {fn } be d-Cauchy, i.e., d(fnR(x), fm (x))dµ(x) → 0. Fix > 0. Then there is N such that for m, n ≥ N, d(fm (x), fn (x))dµ(x) < 2 , so µ({x : d(fm (x), fn (x))} > } < . Thus we can find n1 < n2 < . . . such that if Ai = {x : d(fni (x), fni+1 (x)) > 2−i }, then µ(Ai ) < 2−i . Thus, by BorelCantelli, ∃N ∀i ≥ N (x 6∈ Ai ), a.e.(x), i.e., ∃N ∀i ≥ N (d(fni (x), fni+1 (x)) ≤ 2−i ), a.e.(x), so {fni (x)} is d-Cauchy and fni (x) → f (x), a.e., for some ˜ n , f ) → 0 and so ˜ Using Lebesgue Dominated Convergence, d(f f ∈ G. i ˜ n , f ) → 0. d(f 2 Thus for each Polish group G, we have a canonical way of enlarging G to ˜ in which G is a closed subgroup. An interesting property a Polish group G of this extension is the following. ˜ is contractible. Proposition 19.7. For any Polish group G, G
128
III. COCYCLES AND COHOMOLOGY
˜ t ∈ [0, 1] let ft (x) = Proof. We can take (X, µ) = ([0, 1], λ). For f ∈ G, ˜ s , gt ) ≤ f (x), if x > t; = 1, if x ≤ t. Thus f0 = f, f1 = 1. Also if s < t, d(f ˜ ˜ to 1. 2 (t − s) + d(f, g), so (f, t) 7→ ft is continuous and a contraction of G We next note that we can extend the definition of L(X, µ, G) to the case of a σ-finite measure. So let X be a standard Borel space, ν a non-atomic σ-finite measure on X and let µ ∼ ν be an equivalent probability measure. Then we define L(X, ν, G) to be L(X, µ, G) (which clearly only depends on ν). The topology τX,µ,G depends only on the measure class of µ, thus only on ν, and we can denote it by τX,ν,G . We also have the compatible metric dX,µ,G , which of course depends on µ. Fix now a complete metric d on G. Consider the complete metric group ˜ = (L(X, µ, G), dX,µ,G ). Then we have a canonical shift action of ˜ d) (G, ˜ which is an analog of the ˜ d), Aut(X, µ) by isometric isomorphisms of (G, 2 Koopman representation of Aut(X, µ) on L (X, µ). Namely, for any T ∈ Aut(X, µ), f ∈ L(X, µ, G), let T · f (x) = f (T −1 (x)). Thus to each T ∈ Aut(X, µ) we associate the isometric isomorphism ιT (f ) = f ◦ T −1 ˜ Endowing, as usual, the isometry group Iso(G, ˜ with the point˜ d) ˜ d). on (G, wise convergence topology, under which it is a Polish group, we have the following fact. Proposition 19.8. If G is non-trivial, the map T 7→ ιT is a topological group isomorphism of (Aut(X, µ), w) with a (necessarily closed) subgroup ˜ In particular, the action T · f of Aut(X, µ) on L(X, µ, G) is ˜ d). of Iso(G, continuous. Proof. This map is clearly a homomorphism from Aut(X, µ) into the ˜ If ιT = 1, then f (T −1 (x)) = f (x), a.e., ∀f ∈ G. ˜ d). ˜ Fix group Iso(G, g0 6= 1, g0 ∈ G. Let for A ⊆ X, χA (x) = g0 , if x ∈ A; = 1, if x 6∈ A. If A is Borel, then χA (T −1 (x)) = χT (A) (x), a.e, so T (A) = A. Thus T = 1. Clearly T 7→ ιT is Borel, thus continuous. Conversely, if ιTn → 1, R ˜ then for any Borel A, ι (χ ) → χ (in G) or d(χ Tn A A Tn (A) (x), χA (x))dµ(x) = R χTn (A)∆A (x)d(1, g0 )dµ(x) = µ(Tn (A)∆A)d(1, g0 ) → 0, i.e., Tn → 1. 2 ˜ Glasner [Gl1] has Remark. It is clear that if G is abelian, so is G. ˜ shown that T is monothetic. Remark. Pestov [Pe1] has shown that if G is amenable locally compact, ˜ is extremely amenable. then G Remark. Let G = Z2 and d be the discrete metric on G. Then clearly ˜ B) = dµ (A, B) = µ(A∆B). ˜ can be identified with (MALGµ , ∆) and d(A, G In this case, via T 7→ ιT defined above, Aut(X, µ) is identified with the group of all isometries S of (MALGµ , dµ ) such that S(∅) = ∅.
20. COCYCLES
129
˜ by shift, T · f = f ◦ T −1 , we can form the Since Aut(X, µ) acts on G semidirect product ˜ Aut(X, µ) n G, ˜ with multiplication which is the product space Aut(X, µ) × G (S, f )(T, g) = (ST, (T −1 · f )g). See Appendix I for more about semidirect products and, in particular, see part (B) for our choice of the multiplication rule. ˜ is a Polish group (when Aut(X, µ) is equipped Clearly Aut(X, µ) n G with the weak topology) and the canonical action of Aut(X, µ) on the group ˜ given by Aut(X, µ) n G T · (S, f ) = (T ST −1 , T · f ) is a continuous action by automorphisms. Comments. The basic facts about the group L(X, µ, G) can be found in Moore [Mo], which lays the foundation of Moore cohomology theory (note that one can also consider the analogous metric space L(X, µ, M ) for any given metric space M ). The earlier paper Hartman-Mycielski [HM] gives a version of this construction with emphasis on the connectedness properties of the resulting groups. 20. Cocycles (A) Let a ∈ A(Γ, X, µ) be a measure preserving action of a countable group Γ on (X, µ) and let G be a Polish group. A cocycle of a (with values in G) is a Borel map α : Γ × X → G such that for all γ, δ ∈ Γ, writing a(γ, x) = γ · x, we have the cocycle identity: (*)
α(γδ, x) = α(γ, δ · x)α(δ, x), µ-a.e.(x).
We identify two cocycles α, β if for all γ, α(γ, x) = β(γ, x), a.e.(x). Thus we could equivalently consider measurable instead of Borel cocycles. Note that the cocycle identity implies α(1, x) = 1, α(γ, x)−1 = α(γ −1 , γ · x), a.e.(x). We denote by Z 1 (a, G) the set of cocycles for the action a. Thus Z 1 (a, G) ⊆ L(Γ × X, ηΓ × µ, G), where ηΓ is the counting measure on Γ. When Γ acts in a Borel way on a standard Borel space X a strict cocycle for this action is a Borel map α : Γ × X → G such that α(γδ, x) = α(γ, δ · x)α(δ, x) for all γ, δ ∈ Γ, and all x ∈ X. It is easy to check that given α ∈ Z 1 (a, G), there is a strict cocycle α0 such that for all γ, α(γ, x) = α0 (γ, x), a.e.(x). Indeed, there is a Borel set A ⊆ X of measure 1 which is invariant under the action a and (∗ ) holds for all γ, δ ∈ Γ, x ∈ A. Then put α0 (γ, x) = α(γ, x), if x ∈ A, and α0 (γ, x) = 1, if x 6∈ A.
130
III. COCYCLES AND COHOMOLOGY
When α ∈ Z 1 (a, G) is independent of x, i.e., α(γ, x) = ϕ(γ),, where ϕ : Γ → G, then (∗ ) implies that ϕ is a homomorphism. Thus the homomorphisms ϕ : Γ → G are exactly the cocycles that depend only on γ ∈ Γ. ˜ acts on Z 1 (a, G) by The group L(X, µ, G) = G f · α(γ, x) = f (γ · x)α(γ, x)f (x)−1 . We say that α, β are cohomologous cocycles or equivalent cocycles, in symbols α ∼ β, ˜ with f · α = β. We denote by if there is f ∈ G [α]∼ the cohomology class of α. The trivial cocycle is denoted by 1: α(γ, x) = 1. ˜ A cocycle α is a coboundary if it is cohomologous to 1, i.e., for some f ∈ G α(γ, x) = f (γ · x)f (x)−1 . The set of coboundaries is denoted by B 1 (a, G). Finally the quotient space H 1 (a, G) = Z 1 (a, G)/ ∼ is called the (1st-)cohomology space of a, relative to G. In the special case where G is abelian, Z 1 (a, G) is an abelian group under pointwise multiplication, B 1 (a, G) is a subgroup and clearly ∼ is induced by the cosets of B 1 (a, G), so H 1 (a, G) = Z 1 (a, G)/B 1 (a, G) is an abelian group, the (1st-)cohomology group of a. If we view an action a ∈ A(Γ, X, µ) as a homomorphism a ∈ Hom(Γ, Aut(X, µ)), where (by abusing notation) a(γ) = γ a = (x 7→ a(γ, x)) and a cocycle α : Γ × X → G as a map α : Γ → L(X, µ, G), where α(γ)(x) = α(γ, x), then, as explained in Appendix I, the pair (a, α) gives an extension of the homomorphism a ∈ Hom(Γ, Aut(X, µ)) to a homomorphism γ 7→ (a(γ), α(γ)) of ˜ where this semidirect product is defined as in Section Γ into Aut(X, µ) n G, 19, (B). Thus if we put AZ 1 (Γ, X, µ, G) = {(a, α) : a ∈ A(Γ, X, µ), α ∈ Z 1 (a, G)}, ˜ then we can identify AZ 1 (Γ, X, µ, G) with Hom(Γ, Aut(X, µ) n G). ˜ by Recall that Aut(X, µ) acts on Aut(X, µ) n G T · (S, f ) = (T ST −1 , T · f ), where T · f = f ◦ T −1 , and thus acts on AZ 1 (Γ, X, µ, G) by T · (a, α) = (T aT −1 , T · α). where T · α(γ, x) = α(γ, T −1 (x)).
20. COCYCLES
131
Let Ca be the stabilizer of the action a, i.e., Ca = {T : T aT −1 = a}. This is a closed subgroup of Aut(X, µ). Then Ca acts on Z1 (a, G) via (T, α) 7→ T · α. Note that this action preserves ∼, i.e., α∼β ⇒T ·α∼T ·β and so preserves B 1 (a, G). It also induces an action on H 1 (a, G). ˜ (a closed subgroup of the Consider now the semidirect product Ca n G ˜ It acts on Z 1 (a, G) by group Aut(X, µ) n G). (T, g) · α(γ, x) = g(γ · T −1 (x))α(γ, T −1 (x))g(T −1 (x))−1 . If α, β belong to the same orbit of this action then we say that α, β are weakly equivalent, in symbols, α ∼w β. We can easily see that α ∼w β ↔ ∃T ∈ Ca (T · α ∼ β). More generically, if ai ∈ A(Γ, Xi , µi ), i = 1, 2, αi ∈ Z 1 (ai , G), we say that α1 , α2 are weakly equivalent, α1 ∼w α2 , if there is an isomorphism of the actions a1 , a2 that sends α1 to a cocycle equivalent to α2 . (B) Let now E be a countable, measure preserving equivalence relation on (X, µ). Recall from Section 6, (B)R that we can defineR a σ-finite Borel measure M on E ⊆ X 2 by M (A) = card(Ax )dµ(x) = card(Ay )dµ(y), for Borel A ⊆ E. Note that A ⊆ E is M -conull iff there is an E-invariant Borel set B ⊆ X of measure 1 such that for x, y ∈ B, xEy ⇒ (x, y) ∈ A. A cocycle of E with values in G is a Borel map α : E → G which satisfies the cocycle identity α(x, z) = α(y, z)α(x, y), for all xEyEz in some E-invariant Borel subset of X of measure 1. We identify two cocycles α, β if α(x, y) = β(x, y), M −a.e.(x, y). Again instead of Borel we could equivalently consider (M −)measurable cocycles. If α is a cocycle, then α(x, x) = 1 and α(x, y) = α(y, x)−1 , a.e. We denote by Z 1 (E, G) the set of cocycles for E with values in G, so that Z 1 (E, G) ⊆ L(E, M, G). Again for any countable Borel equivalence relation E on a standard Borel space X, a strict cocycle is a Borel map α : E → G such that α(x, z) = α(y, z)α(x, y) for all xEyEz. It is easy to see that for any α ∈ Z 1 (E, G) there is a strict cocycle α0 such that α(x, y) = α0 (x, y), M −a.e.(x, y). ˜ acts on Z 1 (E, G) by The group L(X, µ, G) = G f · α(x, y) = f (y)α(x, y)f (x)−1 ,
132
III. COCYCLES AND COHOMOLOGY
and we say that α, β are cohomologous cocycles or equivalent cocycles, in symbols α ∼ β, if they belong to the same orbit of this action and we denote by [α]∼ the cohomology class of α. The cohomology class of the trivial cocycle α(x, y) = 1 is the set of coboundaries, denoted by B 1 (E, G), and the quotient space H 1 (E, G) = Z 1 (E, G)/ ∼ is the (1st-)cohomology space of E, relative to G. Again if G is abelian, B 1 (E, G) ≤ Z 1 (E, G) and H 1 (E, G) = Z 1 (E, G)/B 1 (E, G) are also abelian groups, the latter called the (1st-)cohomology group of E. The automorphism group N [E] of [E] acts on Z 1 (E, G) via T · α(x, y) = α(T −1 (x), T −1 (y)), and this action preserves ∼ and B 1 (E, G) and thus descends to an action on H 1 (E, G). Finally N [E] n L(X, µ, G) ≤ Aut(X, µ) n L(X, µ, G) acts on Z 1 (E, G) by (T, g) · α(x, y) = g(T −1 (y))α(T −1 (x), T −1 (y))g(T −1 (x))−1 and induces the notion of weak equivalence, α ∼w β, among cocycles. Again α ∼w β ⇔ ∃T ∈ N [E](T · α ∼ β). Similarly we define the notion of weak equivalence of αi ∈ Z 1 (Ei , G), i = 1, 2. Remark. Note that if T ∈ [E], then T · α ∼ α. Because if f (x) = α(x, T −1 (x)), then f · α(x, y) = f (y)α(x, y)f (x)−1 = α(y, T −1 (y))α(x, y)α(T −1 (x), x) = α(T −1 (x), T −1 (y)) = T · α(x, y), so α ∼ f · α = T · α. (C) Suppose now that a ∈ A(Γ, X, µ) and let Ea be the associated equivalence relation. Then there is a canonical embedding of Z 1 (Ea , G) into Z 1 (a, G) given by α(x, y) 7→ α∗ (γ, x) = α(x, γ · x). (To see that this is 1-1 note that if α 6= β, so that M ({(x, y) ∈ Ea : α(x, y) 6= β(x, y)}) > 0, then for some γ ∈ Γ, M ({(x, γ · x) : α(x, γ · x) 6= β(x, γ · x)}) > 0, so µ({x : α∗ (γ, x) 6= β ∗ (γ, x)}) > 0.) Clearly α 7→ α∗ preserves the action of L(X, µ, G) on Z 1 (Ea , G), Z 1 (a, G), so sends Z 1 (Ea , G) onto a cohomology invariant subset of Z 1 (a, G). Also it maps B 1 (Ea , G) onto B 1 (a, G).
20. COCYCLES
133
We note that α 7→ α∗ maps Z 1 (Ea , G) onto Z 1 (a, G) iff a is free. Indeed, if a is free, and α ∈ Z 1 (a, G), let β(x, y) = α(γ, x), where γ · x = y. Then β ∗ = α. Conversely, if a is not free let β(γ, x) = γ. If α∗ = β, then γ = α(x, γ · x). Let x be such that γ · x = x for some γ 6= 1. Then γ = α(x, γ · x) = α(x, x) = 1, a contradiction. Remark. Suppose that a is not necessarily free but Ea is treeable with associated treeing T ⊆ Ea . Let Γ = {γn } and find Borel e : T → Γ with e(x, y) · x = y and e(y, x) = e(x, y)−1 . Also let e(x, x) = 1. Extend e to : E → Γ by (x, y) = e(xn , y) · · · e(x1 , x2 )e(x, x1 ), where x, x1 , . . . , xn , y is the unique T -path from x to y. Clearly ∈ Z 1 (Ea , Γ) and (x, y) · x = y. Define now β ∈ Z 1 (a, G) 7→ β∗ ∈ Z 1 (Ea , G) by β∗ (x, y) = β((x, y), x). ∗ Then (α )∗ = α, so β 7→ β∗ is a map from Z 1 (a, G) onto Z 1 (Ea , G), which is the inverse of α 7→ α∗ on Z 1 (Ea , G)∗ . (D) Suppose that {γi }i∈I (finite or infinite) is a set of generators for Γ. Fix a ∈ A(Γ, X, µ). Given α ∈ Z 1 (a, G), put fα,i (x) = α(γi , x). Clearly ˜ Since for δ, δ1 , . . . , δn ∈ Γ, fα,i ∈ L(X, µ, G) = G. α(δ1 δ2 · · · δn , x) = α(δ1 , δ2 · · · δn · x) · · · α(δn−1 , δn · x)α(δn , x), and α(δ −1 , x) = α(δ, δ −1 · x)−1 , it follows that the maps {fα,i }i∈I completely determine α. When Γ = Fn and γ1 , . . . , γn are free generators, then α 7→ (fα,1 , . . . , fα,n ) is a 1-1 correspondence between Z 1 (a, G) and ˜ n , so Z 1 (a, G) can be identified with (G) ˜ n . When G is abelian, this (G) 1 ˜ n. is an identification of the group Z (a, G) with the product group (G) 1 Thus in the particular case Γ = F1 = Z, Z (a, G) can be identified with L(X, µ, G) with α ∈ Z 1 (a, G) identified with f (x) = α(1, x) (then note that α(n, x) = f ((n − 1) · x) . . . f (1 · x)f (x) if n > 0 and α(n, x) = α(−n, n · x)−1 , if n < 0). (E) Given two homomorphisms ϕ, ψ : Γ → G between countable groups Γ, G viewed as cocycles in Z 1 (a, G) for a given a ∈ A(Γ, X, µ), we can determine when ϕ, ψ are cohomologous. Call ϕ, ψ conjugate if there is g ∈ G such that gϕ(γ)g −1 = ψ(γ), ∀γ ∈ Γ. Proposition 20.1. Let Γ be a countable group, G a countable group and ϕ, ψ ∈ Hom(Γ, G). If a ∈ A(Γ, X, µ) is such that every finite index subgroup of Γ acts ergodically (e.g., if a is weak mixing), then (viewing ϕ, ψ as cocycles in Z 1 (a, G)), ϕ ∼ ψ ⇔ ϕ, ψ are conjugate. Proof. ⇐ is obvious. ˜ ⇒: Assume that for some f ∈ G, ψ(γ) = f (γ · x)ϕ(γ)f (x)−1 .
134
III. COCYCLES AND COHOMOLOGY
Let g0 ∈ G be such that f −1 ({g0 }) = A0 has positive measure. Then x, γ · x ∈ A0 ⇒ ψ(γ) = g0 ϕ(γ)g0−1 . Put Γ0 = {γ ∈ Γ : ψ(γ) = g0 ϕ(γ)g0−1 }. Then Γ0 ≤ Γ and x, γ · x ∈ A0 ⇒ γ ∈ Γ0 . Conversely, if γ ∈ Γ0 , x ∈ A0 , then ψ(γ) = f (γ · x)ϕ(γ)g0−1 , so g0 ϕ(γ)g0−1 = f (γ · x)ϕ(γ)g0−1 , thus f (γ · x) = g0 and so γ · x ∈ A0 . It follows that A0 is Γ0 -invariant and A0 meets every Γ-orbit in a single Γ0 orbit (modulo null sets). Let {γi }i∈I be a set of left-coset representatives for Γ0 . Then we claim that i 6= j ⇒ γi · A0 ∩ γj · A0 = ∅. Otherwise, there are x0 , y0 ∈ A0 with γi · x0 = γj · y0 , so x0 , y0 are in the same Γ-orbit, thus the same Γ0 -orbit. Let γ0 · x0 = y0 , for some γ0 ∈ Γ0 . Then γ0−1 γj−1 γi · x0 = x0 , so γ0−1 γj−1 γi ∈ Γ0 , i.e., γi ∈ γj Γ0 , a contradiction. Therefore I is finite and thus [Γ : Γ0 ] < ∞. So Γ0 acts ergodically and therefore µ(A0 ) = 1 and ϕ, ψ are conjugate. 2 In particular, under the hypotheses of the proposition, the only homomorphism in B 1 (a, G) is the trivial one. Remark. The condition that every finite index subgroup of Γ acts ergodically is necessary. Let Γ = Γ0 × Z2 , where Γ0 is an infinite group. Let a ∈ A(Γ0 , X, µ) be free and ergodic, let Y = X × {0} ∪ X × {1} = X0 ∪ X1 be the union of two disjoint copies of X, and let ν = 21 (µ0 + µ1 ), with µi the copy of µ in Xi . Let Γ0 act on Y in the obvious way: γ · (x, i) = (γ · x, i). Let Z2 act on Y by the involution T (x, i) = (x, 1−i). This action commutes with the Γ0 -action, so it gives an action of Γ which is free and ergodic. However, it is obviously not Γ0 -ergodic. Let ϕ : Γ → Z2 be the projection map. Then ϕ ∼ 1 (since ϕ(γ) = f (γ · y) − f (y), where f (x, i) = i) but ϕ 6= 1. (F) Let E1 ⊆ E2 be two equivalence relations on (X, µ). Then there is a canonical restriction map α ∈ Z 1 (E2 , G) 7→ α|E1 ∈ Z 1 (E1 , G). ˜ and thus α ∼ β ⇒ α|E1 ∼ β|E1 . It clearly respects the action of G (G) Consider next an equivalence relation E on (X, µ) and let A ⊆ X be a Borel complete section a.e. (i.e., A meets almost all E-classes). Then E|A is an equivalence relation on (A, µA ), where µA (B) = µ(B) µ(A) is the normalized 1 restriction of µ to A. For each α ∈ Z (E, G), we denote by α|A its restriction to A, so that α|A ∈ Z 1 (E|A, G). Conversely, fix a Borel map fA : X → A such that fA (x)Ex, a.e., and fA |A = id. Then given β ∈ Z 1 (E|A, G), define β A ∈ Z 1 (E, G), by β A (x, y) = β(fA (x), fA (y)). Thus β A |A = β. It is clear that β1 ∼ β2 ⇔ β1A ∼ β2A . We also claim that α1 ∼ α2 ⇔ α1 |A ∼ α2 |A.
21. THE MACKEY ACTION AND REDUCTION COCYCLES
135
The direction ⇒ is obvious. Conversely, assume that there is g : A → G such that for uEv, u, v ∈ A, we have α2 (u, v) = g(v)α1 (u, v)g(u)−1 . Let then f : X → G be defined by f (x) = α2 (fA (x), x)g(fA (x))α1 (x, fA (x)). Then for xEy we have f (y)α1 (x, y)f (x)−1 = α2 (fA (y), y)g(fA (y))α1 (y, fA (y))α1 (x, y) α1 (fA (x), x)g(fA (x))−1 α2 (x, fA (x)) = α2 (fA (y), y)g(fA (y))α1 (fA (x), fA (y))g(fA (x))−1 α2 (x, fA (x)) = α2 (fA (y), y)α2 (fA (x), fA (y))α2 (x, fA (x)) = α2 (x, y). In particular, for any α ∈ Z 1 (E, G) α ∼ (α|A)A (since α|A = (α|A)A |A). Thus the map β ∈ Z 1 (E|A, G) 7→ β A ∈ Z 1 (E, G) 1 ⊆ Z 1 (E, G), which is a is a bijection from Z 1 (E|A, G) onto a subset ZA 1 complete section of ∼ on Z (E, G). So in a sense the cohomology spaces H 1 (E, G) and H 1 (E|A, G) are isomorphic.
Comments. Some basic references for the theory of cocycles are the following: Feldman-Moore [FM], Schmidt [Sc1], [Sc5] and Zimmer [Zi]. 21. The Mackey action and reduction cocycles (A) We will work for a while in a pure Borel theoretic context with no measures present. Let E be a countable Borel equivalence relation on a standard Borel space X, let G be a Polish group, and α : E → G a strict Borel cocycle. Consider the space X × G and the equivalence relation Rα on X × G defined by (x, g)Rα (y, h) ⇔ xEy & α(x, y)g = h. Let Xα = (X × G)/Rα be the quotient space, which is not necessarily a standard Borel space. The group G acts in a Borel way on X × G by g · (x, h) = (x, hg −1 ) and this action respects Rα , so G acts on Xα by g · [x, h] = [x, hg −1 ],
136
III. COCYCLES AND COHOMOLOGY
where [x, h] = [(x, h)]Rα . Note that this action is free. Y the Recall that when a group H acts on a set Y , we denote by EH equivalence relation induced by this action. Let pα : X → Xα be defined by pα (x) = [x, 1]. Xα Then it can be easily seen that pα is a reduction of E to EG , i.e., Xα xEy ⇔ pα (x)EG pα (y),
and α is the associated cocycle to pα in the sense that xEy ⇒ α(x, y) · pα (x) = pα (y). Moreover the reduction pα admits a Borel lifting, i.e., there is a Borel map ϕα : X → X × G (namely ϕα (x) = (x, 1)) with pα (x) = [ϕα (x)]. Conversely, let Y be a standard Borel space, F a Borel equivalence relation on Y , G a Polish group and assume G acts freely on Y /F so that this action has a Borel lifting, in the sense that there is a Borel map ψ : G×Y → Y /F Y such that g·[y]F = [ψ(g, y)]F . If p : X → Y /F is a reduction of E to EG , Y /F i.e., xEy ⇔ p(x)EG p(y), and p admits a Borel lifting, i.e., there is Borel ϕ : X → Y with [ϕ(x)]F = p(x), then we can define the (strict) Borel cocycle α : E → G associated to this reduction by α(x, y)·p(x) = p(y) (to see that α is Borel notice that α(x, y) = g ⇔ g · [ϕ(x)]F = [ϕ(y)]F ⇔ ψ(g, ϕ(x))F ϕ(y), so the graph of α is Borel). Let us note now that the map pα : X → Xα has range pα (X) which is a Xα , i.e., G · pα (X) = Xα . We now have the following complete section for EG uniqueness property. Proposition 21.1. Let E be a countable Borel equivalence relation, G a Polish group and α : E → G a strict Borel cocycle. Let Y be a standard Borel space, F a Borel equivalence relation on Y and assume G acts freely Y /F on Y /F with Borel lifting and p : X → Y /F is a reduction of E to EG with Borel lifting, such that G · p(X) = Y /F and p has associated cocycle α. Then there is a unique isomorphism π : Xα → Y /F of G-actions such that π◦pα = p. Moreover, π has a Borel lifting (i.e., there is Borel θ : X ×G → Y with π([x, g]) = [θ(x, g)]F ). Proof. To prove uniqueness, assume that π1 , π2 are two such maps. Then π1 ([x, 1]) = p(x) = π2 ([x, 1]), so π1 ([x, g]) = π1 (g −1 · [x, 1]) = g −1 · π1 ([x, 1]) = g −1 · π2 ([x, 1]) = π2 ([x, g]). Now put π([x, g]) = g −1 · p(x). It is easy to check that this is well-defined and works. Let now α ∼ β in the Borel sense, i.e., there is Borel f : X → G with f (y)α(x, y)f (x)−1 = β(x, y).
2
21. THE MACKEY ACTION AND REDUCTION COCYCLES Y /F
Let α come from a reduction p of E to EG
137
as in 21.1. Then let
q(x) = f (x) · p(x), Y /F
q : X → Y /F . Then q is a reduction with Borel lifting of E to EG with asY /F sociated cocycle β(x, y). Moreover q(x)EG p(x). Clearly (Xα , pα ; Y /F, p), (Xβ , pβ ; Y /F, q) are canonically isomorphic by 21.1 and chasing diagrams this shows that Xα , Xβ are isomorphic G-spaces via the isomorphism ρ : Xα → Xβ given by ρ([x, g]Rα ) = [x, f (x)g]Rβ Xβ ρ(pα (x))EG pβ (x). Conversely, if ρ
: Xα → Xβ is an isomorand moreover X phism of G-actions with Borel lifting and ρ(pα (x))EG β pβ (x) with ρ(pα (x)) = f (x)−1 · pβ (x), then f is Borel and f (y)α(x, y)f (x)−1 = β(x, y), so α ∼ β, in the Borel sense. Thus there is a canonical 1-1 correspondence between f : X → G as above, witnessing that α ∼ β, in the Borel sense, and isomorphisms of Gactions ρ : Xα → Xβ with Borel liftings for which X
ρ(pα (x))EG β pβ (x). The action of G on the space Xα is called the Mackey action (or Mackey range or Poincar´e flow ) associated with the cocycle α. By the above it is an invariant of cohomology (in the pure Borel context). If a countable group Γ acts in a Borel way on X and α : Γ × X → G is a (strict) Borel cocycle for this action, then one can define the skew product action of Γ on X × G given by γ · (x, g) = (γ · x, α(γ, x)g). Denoting by Rα the induced equivalence relation, we can define the Mackey action as before and prove similar facts. In particular, if E is a countable Borel equivalence relation induced by a Borel action of a countable group Γ on X, E = EΓX , and α : E → G is a Borel cocycle, then α∗ (γ, x) = α(x, γ · x) is a Borel cocycle for the action and it is easy to check that Rα∗ = Rα , so Rα is induced by the skew product action of Γ on X × G associated to α∗ . (B) We are of course primarily interested in the case where the space Xα is standard, i.e., where the countable Borel equivalence relation Rα is smooth (equivalently admits a Borel selector). Let us introduce the following notions. For each Borel cocycle α : E → G the kernel of α is the equivalence relation xEα y ⇔ xEy & α(x, y) = 1.
138
III. COCYCLES AND COHOMOLOGY
We say that α is smooth if Eα is smooth. A Borel cocycle α : E → G is called a reduction cocycle if there is a free Borel action of G on a standard Borel space Y and a Borel reduction Y with associated cocycle α, α(x, y) · p(x) = p(y). p : X → Y of E to EG More generally, if G, Y are as before and p : X → Y is only a Borel Y , i.e., homomorphism of E to EG Y xEy ⇒ p(x)EG p(y),
then p has an associated cocycle α : E → G given by α(x, y) · p(x) = p(y). Again if α ∼ β with β(x, y) = f (y)α(x, y)f (x)−1 , where f : X → G is Borel, then letting q(x) = f (x) · p(x), we have that q is a Borel homomorphism of Y such that q(x)E Y p(x) and q has associated cocycle β. E to EG G We now have the following fact: Proposition 21.2. Let E be a countable Borel equivalence relation, G a countable group and α : E → G a Borel cocycle. Then the following are equivalent: (i) α is smooth, (ii) Rα is smooth, (iii) α is a reduction cocycle, (iv) α is associated to a countable-to-1 Borel homomorphism. Proof. The previous analysis of the Mackey action shows that (ii) ⇒ (iii) and (iii) ⇒ (i) is clear as if α is associated to the reduction p, then xEα y ⇔ p(x) = p(y). Clearly (iii) ⇒ (iv) and we can see that (iv) ⇒ (i) as follows: If p is a countable-to-1 homomorphism and xEy ⇒ α(x, y) · p(x) = p(y), then Eα ⊆ {(x, y) : p(x) = p(y)} = F , which is a smooth countable Borel equivalence relation, thus Eα is smooth. Finally, we verify that (i) ⇒ (ii): Let c = [x, g]Rα . Let gc = the least g 0 such that ∃y(y, g 0 ) ∈ c (“least” means in some fixed enumeration of G). Then if (y, gc ), (z, gc ) ∈ c, we have yEα z. There is Borel f : X × G → X × G such that f (x, g) = (y, gc ), where c = [x, g]Rα , (y, gc ) ∈ c. Then (x, g)Rα (y, h) ⇔ f (x, g)Eα × ∆(G)f (y, h), where ∆(G) is the equality relation on G. Thus Rα is smooth. 2 A special case of a reduction cocycle comes from isomorphism. Let G be a countable group. A Borel cocycle α : E → G is called an isomorphism cocycle if there is a free Borel action of G on a standard Borel space Y Y with associated cocycle and a Borel isomorphism p : X → Y of E with EG α, α(x, y) · p(x) = p(y). In this case α has the following property: (*) For each x ∈ X, y 7→ α(x, y) is a bijection of [x]E with G. Conversely, if α satisfies (*), then it is easy to check that it is an isomorphism cocycle. Indeed, notice that (*) implies that α(x, y) = 1 ⇔ x = y, so α is smooth and thus the space Xα of the Mackey action of G is standard Borel, and of course the action is free. Finally (*) implies that the canonical Xα map pα (x) = [x, 1] is a Borel isomorphism of E with EG .
21. THE MACKEY ACTION AND REDUCTION COCYCLES
139
(C) We now move on to the measure theoretic framework. Let E be a countable Borel, measure preserving equivalence relation on (X, µ), G a locally compact Polish group, ηG a left-invariant Haar measure, and µG ∼ ηG an equivalent probability measure on G. Given α ∈ Z 1 (E, G), which we can assume to be strict, consider the product space (X × G, µ × µG ) and the equivalence relation Rα given as before by (x, g)Rα (y, h) ⇔ xEy & α(x, y)g = h. It is easy to see that Rα is (µ × ηG )−measure preserving and therefore it is (µ × µG )−quasi-invariant, i.e., Rα -saturations of null sets are null. Now let Sα be the σ-subalgebra of MALGµ×µG which consists of the Rα invariant sets. Then there is (an essentially unique) standard Borel space Xα and a Borel map πα : X ×G → Xα , which is Rα -invariant, and such that if (πα )∗ (µ × µG ) = µα , then A 7→ (πα )−1 (A) is an isomorphism of MALGµα with Sα (see, e.g., Kechris [Kec2], 17.43). Now G acts on X × G by g · (x, h) = (x, hg −1 ) and µ × µG is quasi-invariant under this action, which also preserves Rα and thus Sα . Then by a result of Mackey [M] (see also Kechris [Kec1], 3.47), there is (an essentially unique) Borel action of G on Xα such that µα is quasiinvariant and πα : X × G → Xα preserves the G-actions. The G-action on (Xα , µα ) is called the Mackey action (or Mackey range or Poincar´e flow ) associated to α (in the measure theoretic context). Define now pα : X → Xα by pα (x) = πα (x, 1). Xα pα (y) pα (x)EG
& α(x, y) · pα (x) = pα (y). However the action Then xEy ⇒ of G on Xα might not be free (in fact Xα can be trivial), and so pα might not be a reduction. If α ∼ β and f : X → G is such that f (y)α(x, y)f (x)−1 = β(x, y), let F : X × G → X × G be defined by F (x, g) = (x, f (x)g). Then F is a Borel isomorphism of Rα with Rβ and also of the G-action and moreover preserves µ × ηG . It therefore induces an isomorphism of the Mackey actions of α, β. Thus the Mackey action is an invariant of cohomology. One can similarly define the Mackey action associated to an action a ∈ A(Γ, X, µ) and cocycle α ∈ Z 1 (a, G). We say that α ∈ Z 1 (E, G) is a reduction cocycle if its restriction to an Einvariant Borel set of measure 1 is a reduction cocycle (in the previous Borel sense). Similarly, we define what it means for α to be associated to a Borel homomorphism. Again if α is a reduction cocycle (resp., associated to a homomorphism) and β ∼ α, then β is a reduction cocycle (resp., associated to a homomorphism). Similarly, we say that α is smooth if its restriction to
140
III. COCYCLES AND COHOMOLOGY
an E-invariant Borel set of measure 1 is smooth (in the Borel sense). Also Rα is smooth if its restriction to a co-null Rα -invariant Borel set is smooth (in the Borel sense). Finally, we recall the following notion of transience for cocycles. The cocycle α ∈ Z 1 (E, G) is transient if there is a set A ⊆ X of positive measure and an open nbhd V of 1 such that if x, y ∈ A, xEy, x 6= y, then α(x, y) 6∈ V . Otherwise α is called recurrent. We now have the following analog of 21.2. Proposition 21.3. The following are equivalent for α ∈ Z 1 (E, G), E ergodic and G countable: (i) α is smooth, (ii) Rα is smooth, (iii) α is a reduction cocycle, (iv) α is associated to a countable-to-1 Borel homomorphism, (v) α is transient. Proof. The equivalence of (i)-(iv) comes from 21.2. (i) ⇒ (v): Let A ⊆ X be a Borel transversal for Eα . Then µ(A) > 0 and if x, y ∈ A, xEy, x 6= y, then α(x, y) 6= 1. (v) ⇒ (ii): Fix A witnessing that α is transient. Then A × {1} ⊆ X × G is a partial transversal for Rα (i.e., meets every Rα class in at most one point) and has positive µ × ηG -measure. Call an Rα -invariant Borel set B ⊆ X × G smooth if Rα |B is smooth. By the countable chain condition there is a largest (modulo null sets) smooth set B0 . As A×{1} ⊆ B0 , clearly (µ × ηG )(B0 ) > 0. Since the action of G preserves Rα , B0 is G-invariant, so has the form B0 = A0 × G, for some A0 ⊆ X of positive measure. Moreover A0 is E-invariant (as x ∈ A0 , yEx ⇒ (x, g)Rα (y, α(x, y)g), so (y, α(x, y)g) ∈ B0 and y ∈ A0 ). By ergodicity of E, µ(A0 ) = 1, so B0 = X × G, modulo null sets, and Rα is smooth. 2 Remark. Note that (in the context of 21.3) no α ∈ B 1 (E, G) is transient, otherwise α would be a reduction cocycle, so 1 ∈ B 1 (E, G) would be a reduction cocycle, thus E would be smooth, a contradiction. In the measure theoretic context, we say that α ∈ Z 1 (E, G), G a countable group, is an isomorphism cocycle if there is a free action of G on a standard Borel space Y and a Borel isomorphism π of E restricted to an invariant Y such that α(x, y) · π(x) = π(y). Such cocycles can conull Borel set with EG again be characterized by the condition that y 7→ α(x, y) is a bijection between [x]E and G, a.e.(x). Similarly, if a ∈ FR(Γ, X, µ), b ∈ FR(G, Y, ν), then an orbit equivalence π between a, b gives rise to a cocycle α as above, which we call an orbit equivalence cocycle. Because cocycles arise canonically from reductions of equivalence relations and, in particular, from orbit equivalences, the study of cocycles has numerous applications in the study of orbit equivalence in ergodic theory as well as in the descriptive theory of Borel equivalence relations; see, for
21. THE MACKEY ACTION AND REDUCTION COCYCLES
141
example, Zimmer [Zi2], Adams-Kechris [AK], Popa [Po2], for a sample of such applications. When E is ergodic, then the Mackey action on (Xα , µα ) is also ergodic, since a G-invariant subset of Xα corresponds to a subset of X × G which is both Rα -invariant and G-invariant and thus either null or co-null. On the other hand the ergodicity of Rα is only true for certain cocycles α. Given a cocycle α ∈ Z 1 (E, G), we define its essential range ER(α) ⊆ G by ER(α) = {g ∈ G : ∀ open nbhd V of g ∀A ⊆ X(µ(A) > 0 ⇒ ∃x, y ∈ A, xEy(α(x, y) ∈ V ))}. Note that if g ∈ ER(α), then for every open nbhd V of g and any A ⊆ X of positive measure, {x ∈ A : ∃y(y ∈ A & xEy & α(x, y) ∈ V )} has actually the same measure as A. It is not hard to see that ER(α) is a closed subgroup of G. First it is clear that 1 ∈ ER(α) and ER(α) is closed under inverses. Let now g, h ∈ ER(α). Fix a countable group Γ and a Borel action of Γ on X with EΓX = E. Fix an open nbhd W of gh and open nbhds U, V of g, h with U V ⊆ W . Let µ(A) > 0. Then we can find γ ∈ Γ, such that if C = {x ∈ A : γ · x ∈ A & α(x, γ · x) ∈ U }, then µ(C) > 0 and of course C ∪ γ · C ⊆ A. Similarly there is δ ∈ Γ, D ⊆ C with µ(D) > 0, D ∪ δ · D ⊆ C and α(x, δ · x) ∈ V, ∀x ∈ D. Then if x ∈ D, y = γδ · x ∈ A and if z = δ · x(∈ C ⊆ A), then α(x, y) = α(z, y)α(x, z) = α(z, γ · z)α(x, δ · x) ∈ U V ⊆ W . So gh ∈ ER(α). That ER(α) is closed is obvious. When G is countable, which is the case we are mostly interested in here, ER(α) is the subgroup of G consisting of all g ∈ G that are in the range of α restricted to every set of positive measure. We now have the following well-known characterization. Proposition 21.4. Let E be ergodic, G countable and α ∈ Z 1 (E, G). Then Rα is ergodic (i.e., the Mackey action is trivial) iff ER(α) = G. Proof. Assume first that Rα is ergodic and fix A ⊆ X with µ(A) > 0 and g ∈ G. Since A × {g}, A × {1} have positive measure, they meet almost every Rα -class, so there are x ∈ A, y ∈ A such that (x, 1)Rα (y, g). Then xEy and α(x, y) = g Conversely, assume that ER(α) = G. We will show that the Rα saturation of any positive measure set is conull. For this it is enough to show that if A ⊆ X, µ(A) > 0 and g0 ∈ G, then for almost all x and any h0 ∈ G we have (x, h0 )Rα (y, g0 ) for some y ∈ A. Recall that for each g ∈ G, {x ∈ A : ∃y ∈ A(xEy & α(x, y) = g)} has the same measure as A, so B = {x : ∀g∃y ∈ A(xEy & α(x, y) = g)} has positive measure and since it is E-invariant, it has measure 1. If x ∈ B, then there is y ∈ A with xEy and α(x, y) = g0 h−1 2 0 , so α(x, y)h0 = g0 and (x, h0 )Rα (y, g0 ). In general, the conjugacy class of ER(α) is not a cohomology invariant (see, e.g., Arnold-Nguyen Dinh Cong-Oseledets [ANO]) but the normal
142
III. COCYCLES AND COHOMOLOGY
subgroup N ER(α) =
\
(gER(α)g −1 )
g∈G
is a cohomology invariant. Indeed, let α ∼ β and assume that ghg −1 ∈ ER(α), ∀g, in order to show that h ∈ ER(β). Fix f such that β(x, y) = f (y)α(x, y)f (x)−1 . Let W be an open set containing h, and find open sets U, V with h ∈ U,S1 ∈ V such that V V −1 U V V −1 ⊆ W . Let {gi } be dense in G, so that i V gi = G. Let µ(A) > 0. Fix then i such that B = A ∩ {x : f (x) ∈ V gi } has positive measure. Since gi−1 hgi ∈ ER(α), we can find x, y ∈ B such that α(x, y) ∈ gi−1 V −1 hV gi and so β(x, y) ∈ V gi gi−1 V −1 hV gi gi−1 V −1 = V V −1 hV V −1 ⊆ W . If we denote, for Polish locally compact G, by G = G∪{∞} the one-point compactification of G, then we let for any cocycle α ∈ Z 1 (E, G), ER(α) = {g ∈ G : ∀ open nbhd V of g ∀A ⊆ X(µ(A) > 0 ⇒ ∃x, y ∈ A, xEy(α(x, y) ∈ V ))}. Thus ER(α) = ER(α) ∩ G. We note now the following fact (see Schmidt [Sc1], 3.15). Proposition 21.5. Let E be ergodic, G countable and α ∈ Z 1 (E, G). If α is transient, then ER(α) = {1, ∞}. In particular, ER(α) = {1}. Proof. Clearly 1 ∈ ER(α). Fix A ⊆ X with µ(A) > 0 and finite K ⊆ G. Then Eα |A is smooth, so let T be a Borel transversal for it. Since E|A is ergodic, each E|A-class contains infinitely many Eα |A-classes. Fix an E|A-class C and let T ∩ C = {x1 , x2 , . . . }, where xi 6= xj if i 6= j. We claim that for some i, α(x1 , xi ) 6∈ K, which shows that ∞ ∈ ER(α). Otherwise, if K = {g1 , . . . , gn }, for some k ≤ n and infinitely many i, α(x1 , xi ) = gk , so there are i 6= j with α(xi , xj ) = 1, i.e., xi Eα xj , a contradiction. Finally, we show that ER(α) = {1}. Otherwise let g ∈ ER(α), g 6= 1. We will show that for every set A ⊆ X of positive measure there are x 6= y ∈ A, xEy and α(x, y) = 1, which implies that α is recurrent, a contradiction. Let Γ be a countable group acting in a Borel way on X so that E = EΓX . Then there is some γ such that B = {x ∈ A : γ · x ∈ A & α(x, γ · x) = g} has positive measure. Note that if x ∈ B, then α(x, γ · x) 6= 1, so γ · x 6= x. Thus we can find Borel C ⊆ B of positive measure such that C ∪ γ · C ⊆ A, C ∩ γ · C = ∅ and α(x, γ · x) = g, ∀x ∈ C. Since g −1 ∈ ER(α), we can find δ ∈ Γ, D ⊆ γ · C of positive measure such that D ∪ δ · D ⊆ γ · C and α(y, δ · y) = g −1 , ∀y ∈ D. Then if x ∈ γ −1 · D ⊆ C (⊆ A), y = γ · x ∈ D, δ · y = δγ · x ∈ A, x 6= δγ · x and α(x, δγ · x) = α(γ · x, δγ · x)α(x, γ · x) = α(y, δ · y)α(x, γ · x) = g −1 g = 1. 2 Next we give some examples of cocycles α with ER(α) = G. Let H be a countable dense subgroup of a compact group K. Let Γ = H ⊕H ⊕. . . . Let ϕ : Γ → G be an epimorphism such that for each n, ϕ|{1}⊕ · · · ⊕ {1} ⊕ H ⊕ H ⊕ . . . , where there are n 1’s, is also an epimorphism.
22. HOMOGENEOUS SPACES AND EFFROS’ THEOREM
143
Consider the compact group K N and the dense subgroup Γ ≤ K N . We let Γ act by left-translation on (K N , µ), where µ is the Haar measure on K N , N so that this action is free, measure-preserving, and ergodic. Let E = EΓK . We will define an α ∈ Z 1 (E, G) with ER(α) = G. If xEy, and γ ∈ Γ is such that γ · x = γx = y, put α(x, y) = ϕ(γ). (Thus if we view α as a cocycle of the Γ-action, α is a homomorphism into G.) We will show that ER(α) = G. Fix g ∈ G and let A ⊆ K N have positive measure. Then AA−1 contains a non-empty nbhd of 1, so for some n, {1} ⊕ · · · ⊕ {1} ⊕ H ⊕ H ⊕ · · · ⊆ AA−1 . Let (h1 , h2 , . . . ) ∈ H ⊕ H ⊕ . . . be such that ϕ(1, 1, . . . 1, h1 , h2 , . . . ) = g. If γ = (1, 1, . . . , 1, h1 , h2 , . . . ), then there are x, y ∈ A with γ = yx−1 , so y = γx = γ · x and α(x, y) = ϕ(γ) = g. For instance, we can take H = Z, K = T and G any countable abelian group. Clearly there is ϕ : Γ → G with the above property. Then E as above is an ergodic hyperfinite equivalence relation and α ∈ Z 1 (E, G) has ER(α) = G. (D) We have seen earlier that examples of cocycles in Z 1 (E, G) arise in a natural way from homomorphisms of the equivalence relation E to equivalence relations induced by free Borel actions of G on a standard Borel space. It is reasonable to ask whether there are cocycles that do not arise in this fashion. We now give a simple example of this sort. Let Γ be a non-amenable countable group that admits an epimorphism ϕ : Γ → G onto a non-trivial abelian group G with infinite kernel. Let a ∈ A(Γ, X, µ) be free, mixing and E0 -ergodic and let E = Ea . We will show that there is a cocycle α ∈ Z 1 (E, G) such that α is not associated Y , where G acts in a to any Borel homomorphism p : X → Y of E to EG Borel way freely on Y . First note that the cocycle α(x, y) = ϕ(γ), where γ · x = y, is not a coboundary. Otherwise α(x, y) = f (y)f (x)−1 , for some f : X → G, so that on a set of positive measure A ⊆ X, f |A is constant. Then x, y ∈ A, γ ·x = y ⇒ γ ∈ ker(ϕ) = Γ0 . So the intersection of Γ0 ·A with any Γ-orbit is a single Γ0 -orbit. Since the action is Γ0 -ergodic, Γ0 · A = X and thus Γ = Γ0 , a contradiction. We now argue that α cannot be associated Y as above. Otherwise, xEy ⇒ p(y) = α(x, y) · p(x) and since E is to a p, EG Y y , a.e.(x). E0 -ergodic, and G is abelian, there is y0 ∈ Y such that p(x)EG 0 So p(x) = f (x) · y0 , for some Borel f : X → G. Then α(x, y) = f (y)f (x)−1 , a contradiction. 22. Homogeneous spaces and Effros’ Theorem (A) Before we continue with the theory of cocycles, in this and the next section we will discuss some general facts concerning continuous actions, that we will need later on. We review first some basic results about homogeneous spaces and discuss Effros’ Theorem. Let G be a Polish group and H ≤ G a closed subgroup. We endow G/H with the quotient topology, under which it is a Polish space. If d is a
144
III. COCYCLES AND COHOMOLOGY
right-invariant compatible metric on G, then d(xH, yH) = inf{d(u, v) : u ∈ xH, v ∈ yH} = inf{d(x, v) : v ∈ yH} is a compatible metric for G/H. The group G acts on G/H by g ·xH = gxH and this action is continuous. If d is (two-sided) invariant, then G acts on G/H by isometries. Moreover d on G/H is complete. To see this, suppose d(xn H, xm H) → 0. Then find n1 < n2 < . . . with d(xni H, xni+1 H) = d(xni , xni+1 H) < 2−i . We can then find h1 , h2 , · · · ∈ H with d(xn1 , xn2 h1 ) < 1/2, d(xn2 , xn3 h2 ) < 1/4, . . . , so d(xn1 , xn2 h1 ) < 1/2, d(xn2 h1 , xn3 h2 h1 ) < 1/4, . . . and the sequence {xn1 , xn2 h1 , xn3 h2 h1 , . . . } is d-Cauchy in G. Since d is complete (being invariant), there is f ∈ G with xni hi−1 hi−2 . . . h1 → f , so d(f, xni H) → 0 and thus d(xni H, f H) → 0, i.e., xni H → f H in G/H and so {xn H} converges. (B) We recall Effros’ Theorem (see Effros [Ef2], Kechris [Kec1] and Becker-Kechris [BK]). Theorem 22.1 (Effros). Let G be a Polish group acting continuously on a Polish space P . For each p ∈ P , let Gp = {g ∈ G : g · p = p} be the stabilizer of p. (a) The following are equivalent for any p ∈ P : (i) The canonical map gGp 7→ g · p is a homeomorphism of G/Gp onto G · p (equivalently g 7→ g · p is open from G onto G · p), (ii) G · p is not meager in its relative topology, (iii) G · p is Gδ in P . P is F (in P 2 ), then E P is smooth iff every G-orbit (b) If moreover EG σ G is locally closed (i.e., the difference of two closed sets or equivalently open in its closure). Corollary 22.2 (Marker, Sami). In the context of 22.1 (a), if G · p is nonmeager in P , then G · p is Gδ . Below we use the following notation: ∀∞ nR(n) ⇔ ∃N ∀n ≥ N R(n) ∃∞ nR(n) ⇔ ∀n∃n ≥ N R(n). Corollary 22.3. If a countable group G acts by homeomorphisms on a Polish space P , then the following are equivalent: P is smooth, (i) EG (ii) ∀p ∈ P ∀{gn } ∈ GN [gn · p → p ⇒ ∀∞ n(gn · p = p)], (iii) ∀p ∈ P ∀{gn } ∈ GN (gn · p → p ⇒ ∃n(gn · p = p)].
23. ISOMETRIC ACTIONS
145
P is Proof. The equivalence of (ii), (iii) is clear. Now we have, since EG Fσ and G is countable: P EG is smooth ⇔ ∀p(G · p is locally closed)
⇔ ∀p(p is isolated in G · p) ⇔ ∀p(p is isolated in G · p) ⇔ ∀p∀{gn } ∈ GN (gn · p → p ⇒ ∀∞ n(gn · p = p)}. 2 Since a point p ∈ P is called recurrent if for every open V 3 p, there P is not smooth iff there is a is g ∈ G with p 6= g · p ∈ V , 22.3 says that EG recurrent point. 23. Isometric actions We will discuss here some general results concerning isometric actions. Suppose G is a Polish group, (M, ρ) a Polish metric space (i.e., ρ is a complete separable metric) and (g, x) 7→ g · x a continuous action of G by isometries on (M, ρ). As with any continuous action of a Polish group on a Polish space, we consider the topological ergodic decomposition of this action induced by the Gδ equivalence relation. x≈y ⇔G·x=G·y and the corresponding Gδ partial ordering x 4 y ⇔ x ∈ G · y ⇔ G · x ⊆ G · y. Proposition 23.1. The partial order x 4 y is an equivalence relation and thus the equivalence classes of the topological ergodic decomposition are closed and consist of the orbit closures. Proof. Assume x 4 y and let gn ∈ G be such that gn · y → x. Then ρ(gn · y, x) → 0, so ρ(y, gn−1 · x) → 0 and y 4 x. 2 Recall that for A, B ⊆ M , ρ(A, B) = inf{ρ(x, y) : x ∈ A, y ∈ B). Proposition 23.2. If G · x 6= G · y, then ρ(G · x, G · y) > 0. Proof. Otherwise, let gn , hn be such that ρ(gn · x, hn · y) → 0, so · x, y) → 0, thus y 4 x, and so y ≈ x, a contradiction. 2
ρ(h−1 n gn
Proposition 23.3. ≈ is a closed equivalence relation. Proof. Let {Vn } be an open basis for M . Put Wn = G · Vn and (x, y) ∈ R ⇔ ∃m∃n(x ∈ Wm , y ∈ Wn and Wm ∩ Wn = ∅). Clearly R is open in M 2 and we prove that (≈) = (∼ R).
146
III. COCYCLES AND COHOMOLOGY
Clearly if (x, y) ∈ R, then x ≈ y fails. Conversely, if G · x 6= G · y, then ρ(G · x, G · y) > > 0. So there are m, n with x ∈ Vm ⊆ Bρ (x, /2), y ∈ Vn ⊆ Bρ (y, /2). If z ∈ Wm ∩ Wn , then there are g, h with g −1 · z ∈ Vm , h−1 ·z ∈ Vn , so ρ(x, g −1 ·z) < /2, ρ(y, h−1 ·z) < /2, thus ρ(g·x, h·y) < , a contradiction. 2 Proposition 23.4. The quotient space M/ ≈ (with the quotient topology) is Polish with compatible complete metric ρ(G · x, G · y). Proof. Since ρ(G · x, G · y) = ρ(G · x, G · y) = ρ(x, G · y), it follows that it satisfies the triangle inequality, and so 23.2 shows that it is a metric. We next check that it is compatible with the quotient topology. We first verify that for y ∈ M, > 0, {G · x : ρ(G · x, G · y) < } is open in the quotient topology or equivalently that A = {x : ρ(G · x, G · y) < } is open in M . But x ∈ A ⇔ ∃g, h ∈ Gρ(g · x, h · y) < , so this is clear. Conversely, suppose V ⊆ (M/ ≈) is open and G · x ∈ V . We will then find > 0 such that ρ(G · x, G · y) < ⇒ G · y ∈ V . Put V˜ = {y : G · y ∈ V }. Then V˜ is open in M , G-invariant, and x ∈ V˜ . So fix > 0 with Bρ (x, ) ⊆ V˜ . If ρ(G · x, G · y) = ρ(x, G · y) < , there is h ∈ G with ρ(x, h · y) < , so h · y ∈ Bρ (x, ) ⊆ V˜ , thus y ∈ h−1 · V˜ = V˜ , and G · y ∈ V . Finally we show that ρ(G · x, G · y) is complete. Let ρ(G · xn , G · xm ) = ρ(xn , G · xm ) → 0. Then we can find n1 < n2 < . . . with ρ(xni , G · xni+1 ) < 2−i . Choose inductively gni ∈ G as follows: Let gn1 = 1. Then let gn2 be such that ρ(xn1 , gn2 · xn2 ) < 2−1 . As ρ(xn2 , G · xn3 ) < 2−2 , let h ∈ G be such that ρ(xn2 , h · xn3 ) < 2−2 , so ρ(gn2 · xn2 , gn2 h · xn3 ) < 2−2 . Put gn3 = gn2 h, etc. Thus ρ(gni · xni , gni+1 · xni+1 ) < 2−i and so gni · xni → y for some y, thus ρ(G · xni , G · y) → 0, i.e., some subsequence of G · xn converges in ρ and thus so does G · xn . 2 We will now consider the structure of the orbits G · x. Proposition 23.5. Assume that the Polish group G admits an invariant metric. Then if an orbit G · x is Gδ , it is actually closed. In particular, every non-meager orbit is clopen and if there is a comeager orbit the action is transitive. Proof. Let d be an invariant metric on G. Let H = Gx = {g ∈ G : g·x = x}. Since G·x is Gδ , by Effros’ Theorem the canonical map gH 7→ g·x is a homeomorphism of G/H with G · x. Recall from Section 22 that d on G/H (i.e., d(g · H, h · H)) is a complete metric on G/H invariant under the G-action on G/H. Suppose, towards a contradiction, that gn · x → y, where y 6∈ G · x. Claim. There is g such that gn H → gH in G/H. Granting this, gn · x → g · x = y, a contradiction. Proof of the claim. Let > 0 and let V = Bd (H, ) be an open ball around H in G/H. Let U = {g ∈ G : gH ∈ V }. Then U · x is
23. ISOMETRIC ACTIONS
147
an open nbhd of x in G · x, so for some δ > 0, Bρ (x, δ) ∩ G · x ⊆ U · x. Since ρ(gm · x, gn · x) = ρ(gn−1 gm · x, x) → 0, we have some N such that gn−1 gm · x ∈ U · x, for m, n ≥ N . So for such m, n, gn−1 gm H ∈ V , thus d(gn−1 gm H, H) < and so d(gm H, gn H) < . Therefore {gn H} is d-Cauchy in G/H, thus gn H → gH for some g. If G · x is a non-meager orbit, then, by 22.2, G · x is Gδ so closed, and has non-empty interior, thus is also open. 2 Therefore, when G admits an invariant metric, the closures of the orbits G · x come in two varieties: (i) Those that are equal to single closed orbit G · x = G · x. (ii) Those for which G · x contains more than one orbit, i.e, G · x is not closed. Then the action of G on the invariant closed set G · x is minimal and every orbit contained in G · x is meager in G · x (by 22.2 and 23.5). In M |(G · x) is not smooth, in fact E can be Borel reduced that case, clearly EG 0 to it (see Becker-Kechris [BK], 3.4.5). So we subdivide the space M into two parts: SMOOTH = {x ∈ M : G · x is closed}, ROUGH = {x ∈ M : G · x is not closed}. Then SMOOTH, ROUGH are ≈-invariant, i.e., are unions of orbit closures. Proposition 23.6. Assume that the Polish group G admits an invariant M is Borel, SMOOTH and ROUGH are Borel. metric. Then if EG M is Borel, there Proof. Denote by d an invariant metric on G. Since EG is a Borel map x 7→ {gxn }n∈N ∈ GN such that {gxn }n∈N is dense in Gx (see Becker-Kechris [BK], 7.1.2). Let {fn } be dense in G. Then
x ∈ SMOOTH ⇔ (g · x 7→ gGx ) is continuous ⇔ (g · x 7→ gGx ) is continuous at x ⇔ ∀∃δ∀g(ρ(g · x, x) < δ ⇒ d(gGx , Gx ) < ) ⇔ ∀∃δ∀n(ρ(fn · x, x) < δ ⇒ d(fn Gx , Gx ) < ) ⇔ ∀∃δ∀n(ρ(fn · x, x) < δ ⇒ ∃k∃`(d(fn gxk , gx` ) < )), where , δ vary over positive rationals. Thus SMOOTH is Borel.
2
We should finally note that under the assumptions of 23.6, the equivaM |SMOOTH is smooth, since we have that E M |SMOOTH lence relation EG G M |ROUGH is not smooth, thus E Borel reduces = (≈ |SMOOTH), while EG 0 to it.
148
III. COCYCLES AND COHOMOLOGY
24. Topology on the space of cocycles (A) Let G be a Polish group, (X, µ) a standard measure space and consider the semidirect product Aut(X, µ) n L(X, µ, G) as in Section 19. Given a countable discrete group Γ, we can form the topological space (Aut(X, µ) × L(X, µ, G))Γ = Aut(X, µ)Γ × L(X, µ, G)Γ with the product topology. Then AZ 1 (Γ, X, µ, G) = Hom(Γ, Aut(X, µ) n L(X, µ, G)) is a closed subspace and we equip it with the subspace topology. Thus for each a ∈ A(Γ, X, µ) = Hom(Γ, Aut(X, µ)), the space Z 1 (a, G) is the section of AZ 1 (Γ, X, µ, G) corresponding to a and Z 1 (a, G) is thus a closed subspace of L(X, µ, G)Γ and therefore a Polish space. For αn , α ∈ Z 1 (a, G), αn → α ⇔ ∀γ ∈ Γ(αn (γ, x) → α(γ, x) in measure) ⇔ ∀γ ∈ Γ∀{ni }∃{mi }({mi } is a subsequence of {ni } and αmi (γ, x) → α(γ, x) pointwise, a.e.) Z ⇔ ∀γ ∈ Γ( d(αn (γ, x), α(γ, x))dµ(x) → 0), where d ≤ 1 is a compatible metric for G. Note that by a simple diagonal argument, we also have αn → α ⇔ ∀{ni }∃{mi }({mi } is a subsequence of {ni } and ∀γ(αmi (γ, x) → α(γ, x) pointwise, a.e.)). 1 If Γ = {γk }∞ k=1 , then we can define a compatible metric for Z (a, G) by
d˜1 (α, β) = d1Γ,X,µ,G (α, β) Z ∞ X −k d(α(γk , x), β(γk , x))dµ(x). 2 = k=1
Equivalently we can view Z 1 (a, G) as a closed subspace ηΓ ×µ, G), P∞of L(Γ×X, −k where ηΓ is the counting measure on Γ. Since ηΓ ∼ k=1 2 δγk , where δγ = Dirac measure at γ ∈ Γ, the metric on L(Γ × X, ηΓ × µ, G) associated with d is clearly d˜1 . Assume now that E is a countable measure preserving equivalence relation on (X, µ). Then Z 1 (E, G) ⊆ L(E, M, G), where M is the usual measure on E (see Section 6, (B)). Moreover, it is easy to see that Z 1 (E, G) is a closed subspace of L(E, M, G). Indeed, assume that αn ∈ Z 1 (E, G) and αn → f ∈ L(E, M, G). Let An ⊆ X be E-invariant of measure 1 such that for xEyEz in An , αn (x, z) = αn (y, z)αn (x, y). Since αn → f in L(E, M, G), there is a subsequence T {ni } such that αni (x, y) → f (x, y), M a.e. So there is E-invariant A ⊆ n An , with µ(A) = 1, and for xEy in A, αni (x, y) → f (x, y), so that for xEyEz in A, f (x, z) = f (y, z)f (x, y), i.e., f ∈ Z 1 (E, G). We equip Z 1 (E, G) with the relative topology, so that it is again a Polish space. If Γ = {γk }∞ k=1 is a countable group acting in a Borel
24. TOPOLOGY ON THE SPACE OF COCYCLES
149
S way on X with EΓX = E, then E = ∞ k=1 graph(γk ). Let µk be the probability measure on graph(γ ) which is the image of µ under x 7→ (x, γk · x). k P −k Then M ∼ 2 µk , so a compatible metric for Z 1 (E, G) is Z ∞ X d˜1 (α, β) = 2−k d(α(x, γk · x), β(x, γk · x))dµ(x). k=1
In particular, if a ∈ A(Γ, X, µ) and α 7→ α∗ is the canonical embedding of Z 1 (Ea , G) into Z 1 (a, G) (see Section 20, (C)), then it is an isometry onto a closed subspace of Z 1 (a, G). Finally, if G is a Polish abelian group, then Z 1 (a, G), Z 1 (E, G) are Polish groups, in fact Z 1 (a, G) (resp., Z 1 (E, G)) is a closed subgroup of L(Γ × X, ηΓ × µ, G) (resp., L(E, M, G)). (B) Consider now the action of L(X, µ, G) on Z 1 (a, G) as in Section 20, (A). It is easy to check that it is continuous. We also have the following, when G admits an invariant metric. Proposition 24.1. Let G be a Polish group which admits an invariant metric d, Γ = {γk }∞ k=1 a countable group, a ∈ A(Γ, X, µ) and consider the metric d˜1 on Z 1 (a, G). Then L(X, µ, G) acts by isometries on (Z 1 (a, G), d˜1 ). Similarly for Z 1 (E, G). Proof. A trivial calculation using the above formula for d˜1 (α, β). 2 As in Section 20, (A) the Polish group Aut(X, µ) also acts on the space AZ 1 (Γ, X, µ, G) by T · (a, α) = (T aT −1 , T · α), where T · α(γ, x) = α(γ, T −1 (x)). This action is continuous and so is the action of Ca on Z 1 (a, G). For any Polish group G and compatible metric d on G, this is easily an action by isometries on (Z 1 (a, G), d˜1 ). Also Ca n L(X, µ, G) acts continuously on Z 1 (a, G) and this is an action by isometries for d˜1 , when d is invariant. Consider now the action of N [E] on Z 1 (E, G) as in Section 20, (B). Let E be aperiodic and consider N [E] as a Polish group with the topology discussed in Section 6. We claim that this action is also continuous. For that it is enough to show that it is separately continuous. First let αn → α in Z 1 (E, G) and let T ∈ N [E], in order to show that T · αn → T · α. We have that for any {mi } there is a subsequence {ni } with αni (x, y) → α(x, y), M -a.e., so αni (T −1 (x), T −1 (y)) → α(T −1 (x), T −1 (y)), M -a.e., since T × T preserves M . Next assume that Tn → T in N [E] and let α ∈ Z 1 (E, G). We need to show that Tn · α → T · α in Z 1 (E, G), i.e., α(Tn−1 (x), Tn−1 (y)) → α(T −1 (x), T −1 (y)) in the space L(E, M, G). Since T × T preserves M , this is equivalent to α(Tn−1 T (x), Tn−1 T (y)) → α(x, y). Since Tn−1 T → 1 in N [E], it is enough to show, that if Tn → 1 in N [E], then Tn · α → α in Z 1 (E, G). Let Γ act on X so that EΓX = E. Then we need to show that for each γ ∈ Γ, Tn ·α(x, γ ·x) →
150
III. COCYCLES AND COHOMOLOGY
R α(x, γ ·x) in L(X, µ, G), i.e., d(α(Tn−1 (x), Tn−1 (γ ·x)), α(x, γ ·x))dµ(x) → 0. If this fails, then, by going to a subsequence, we can assume that for some > 0 the above integral is > 3 for each n. By 6.2, since Tn → 1 in N [E], we can assume du (γTn−1 , Tn−1 γ) < , i.e., µ({x : γ · Tn−1 (x) 6= Tn−1 (γ · x)}) < R that−1 . Thus d(α(Tn (x), γ·Tn−1 (x)), α(x, γ·x))dµ(x) > . Let f (x) = α(x, γ·x), so that f ∈ L(X, µ, G). Again by 6.2 we have Tn → 1 in Aut(X, µ) with the weak topology and since the action of Aut(X, L(X, µ, G) is continuous, R µ) on −1 we have Tn · f → f in L(X, µ, G), i.e., d(f (Tn (x)), f (x))dµ(x) → 0, a contradiction. Again the action of N [E] on (Z 1 (a, G), d˜1 ) is by isometries for any Polish group G and compatible metric d. Finally N [E] n L(X, µ, G) acts continuously on Z 1 (E, G) and moreover by isometries for d˜1 if d is also invariant. Remark. As we discussed in Section 20, (D), there is a canonical 1-1 correspondence between Z 1 (a, G), where a ∈ A(Fn , X, µ) and L(X, µ, G)n . It is clear that this is a homeomorphism. Remark. Let E1 ⊆ E2 . Then the map α ∈ Z 1 (E2 , G) 7→ α|E1 ∈ 1 , G) is continuous and respects the actions of L(X, µ, G).
Z 1 (E
(C) Finally recall Section 20, (G). Given an equivalence relation E on (X, µ) and a complete section A ⊆ X, the map α ∈ Z 1 (E, G) 7→ α|A ∈ Z 1 (E|A, G) is continuous and the map β ∈ Z 1 (E|A, G) 7→ β A ∈ Z 1 (E, G) is 1 of Z 1 (E, G), which a homeomorphism of Z 1 (E|A, G) with a Gδ subset ZA is a complete section of ∼ on Z 1 (E, G). Since α1 ∼ α2 ⇔ α1 |A ∼ α2 |A and β1 ∼ β2 ⇔ β1A ∼ β2A , the Borel complexity of ∼ on Z 1 (E, G) and ∼ on Z 1 (E|A, G) is the same and similarly the Borel complexity of each cohomology class [α]∼ is the same as that of [α|A]∼ . 25. Cohomology I: Some general facts We will study from now on the structure of the equivalence relation ∼ of cohomology on Z 1 (a, G) and Z 1 (E, G). We are primarily interested in the case where G is countable but occasionally we will formulate statements in greater generality. If the Polish group G admits an invariant metric d, which is necessarily complete, then, as we have seen in Section 24, the Polish group ˜ acts continuously ˜ = L(X, µ, G), which admits the invariant metric d, G 1 1 1 1 ˜ ˜ by isometries on (Z (a, G), d ) and (Z (E, G), d ). So the theory of Section 23 applies. Of course the equivalence relation induced by this action is the relation of cohomology, ∼. We will denote by ROUGH(a, G) (resp., SMOOTH(a, G)) the rough, resp., smooth parts of this action. Similarly for ROUGH(E, G) , SMOOTH(E, G). Z 1 (a,G) In view of 23.6, we will consider conditions under which ∼ (= EG˜ ) is Borel. Proposition 25.1. Let G be a Polish group admitting an invariant metric d. Let a ∈ ERG(Γ, X, µ). Then for any α ∈ Z 1 (a, G), there is a closed
25. COHOMOLOGY I: SOME GENERAL FACTS
151
˜ ˜ α , d) subgroup H ≤ G and an isometric isomorphism of the stabilizer (G 1 with (H, d). Similarly for any ergodic E and α ∈ Z (E, G). In particular, ˜ α. if G is locally compact (resp., countable), so is each stabilizer G ˜ α , we have Proof. For each f, g ∈ G f (γ · x) = α(γ, x)f (x)α(γ, x)−1 , g(γ · x) = α(γ, x)g(x)α(γ, x)−1 . Thus d(f (γ ·x), g(γ ·x)) = d(f (x), g(x)) and x 7→ d(f (x), g(x)) is Γ-invariant, so, by ergodicity, constant a.e. Thus ˜ g), a.e.(x). d(f (x), g(x)) = d(f, ˜ α be a dense subgroup of G ˜ α . Then Let now G0 ⊆ G ˜ g)], a.e.(x). ∀f, g ∈ G0 [d(f (x), g(x)) = d(f, For each g ∈ G0 , recall that g is an equivalence class of Borel functions from X to G modulo null sets. Pick a representative g : X → G for each g ∈ G0 . Then for almost all x, and all f, g ∈ G0 we have ˜ g) d(f (x), g(x)) = d(f, and 1(x) = 1, f g(x) = f (x)g(x), (f −1 )(x) = (f )−1 (x). So fix x0 ∈ X such that all of the above hold and let F : G0 → G be defined by F (g) = g(x0 ). ˜ → (G, d) is an isometric group embedding and thus Then clearly F : (G0 , d) ˜ α onto a closed subgroup uniquely extends to an isometric isomorphism of G H of G. 2 Corollary 25.2 (Hjorth [Hj1]). In the context of 25.1, if G is additionally locally compact (e.g., if G is countable), then the cohomology relation ∼ is Borel. ˜ be Proof. Consider the case of Z 1 (a, G) and let R ⊆ Z 1 (a, G)2 × G defined by R(α, β, f ) ⇔ f · α = β. It is clearly closed and for each α, β, its (α, β)-section Rα,β is a translate of ˜ α , thus locally compact and so Kσ . By the Arsenin-Kunugui Theorem (see G Kechris [Kec2], 35.46) its projection on Z 1 (a, G)2 , which is the cohomology relation ∼, is Borel. 2 We also have: Proposition 25.3. Let G be a Polish abelian group. Let a ∈ A(Γ, X, µ). Then the cohomology relation on Z 1 (a, G) is Borel. Similarly for cohomology of equivalence relations. ˜ · 1 = B 1 (a, G) is a Borel subgroup of the Polish Proof. Note that G 1 group Z (a, G), and ∼ is induced by the cosets of B 1 (a, G), so is Borel. 2
152
III. COCYCLES AND COHOMOLOGY
It now follows from the two preceding results and 23.6 that if G is locally compact admitting an invariant metric or if G is abelian, then the sets SMOOTH(a, G), SMOOTH(E, G) are Borel. ˜ 1 . First notice One can actually obtain more precise information for G ˜ 1 , i.e., f (γ · x) = f (x), and a ∈ ERG(Γ, X, µ), then f is that if f ∈ G ˜ 1 = G (≤ G). ˜ constant. Thus a ∈ ERG(Γ, X, µ) ⇒ G ˜ 1 as follows: In the case that a is not ergodic, we can determine G Consider the ergodic decomposition of a ∈ A(Γ, X, µ), which is given by a Borel a-invariant surjection π : X → E, where E is the standard Borel space of ergodic invariant measures for a, such that for each e ∈ E, π −1 ({e}) = Xe , is a-invariant, e(Xe ) = 1 R(and e is the unique invariant measure for a|Xe ), and if ν = π∗ µ, then µ = edν(e). Note now that the map f 7→ f ◦ π, (L(E, ν, G), dE,ν,G ) → (L(X, µ, G), dX,µ,G ), is an isometric group embedding, as Z dX,µ,G (f ◦ π, g ◦ π) = d(f (π(x)), g(π(x)))dµ(x) Z Z = d(f (π(x)), g(π(x)))de(x) dν(e) Z = d(f (e), g(e))dν(e) = dE,ν,G (f, g). Thus L(E, ν, G) can be identified, via f 7→ f ◦ π, with a closed subgroup of L(X, µ, G). ˜ 1 . Then again f (γ · x) = f (x), so for each e ∈ E, f |Xe is Let now f ∈ G constant, say f (x) = g(e), for x ∈ Xe . Then f = g ◦ π, thus f ∈ L(E, ν, G). It follows that ˜ 1 = L(E, ν, G). G In the special case where G is abelian, the map f ∈ L(X, µ, G) 7→ f · 1 ∈ B 1 (a, G) is a continuous homomorphism of L(X, µ, G) onto B 1 (a, G), so there is a continuous injective homomorphism of the Polish abelian group L(X, µ, G)/L(E, ν, G) onto B 1 (a, G) and thus the latter is a Polishable subgroup of Z 1 (a, G). The same results apply to Z 1 (E, G) by using the ergodic decomposition of E. We finally record the following simple facts concerning the stabilizers ˜ α. G Proposition 25.4. i) Let G be countable, a ∈ A(Γ, X, µ) and α ∈ Z 1 (a, G). ˜ α , then f : X → CG (ER(α)) (where for H ≤ G, CG (H) = the If f ∈ G centralizer of H in G). Similarly for equivalence relations.
26. COHOMOLOGY II: THE HYPERFINITE CASE
153
ii) Let now G be countable and a ∈ WMIX(Γ, X) (or just assume that every subgroup of finite index in Γ acts ergodically). Then for any homo˜ ϕ = CG (ϕ(Γ)). morphism ϕ : Γ → G (viewed as an element of Z 1 (a, G)), G ˜ α , g ∈ ER(α). If h ∈ G is such that A = f −1 ({h}) Proof. i) Let f ∈ G has positive measure, then for some x, y ∈ A, α(x, y) = g, so hgh−1 = g, thus h ∈ CG (ER(α)) and f : X → CG (ER(α)). ˜ ϕ . Then as in the proof of 20.1, f is constant, say ii) Assume f ∈ G ˜ ϕ ≤ CG (ϕ(Γ)) f (x) = f0 a.e. Then clearly ∀γ(f0 ϕ(γ)f0−1 = ϕ(γ)), i.e., G ˜ and therefore Gϕ = CG (ϕ(Γ)). 2 We conclude with the following problem. Problem 25.5. Let a ∈ ERG(Γ, X, µ) and let G be countable. Is the cohomology relation ∼ on Z 1 (a, G) Fσ (as a subset of Z 1 (a, G)2 )? More generally, for G locally compact admitting an invariant metric, what is the Borel complexity of ∼? Similarly for abelian G. Finally, for what a ∈ A(Γ, X, µ) and Polish G is the cohomology relation ∼ on Z 1 (a, G) Borel? Similarly for cohomology of equivalence relations. 26. Cohomology II: The hyperfinite case We start with the following simple proposition. Proposition 26.1. Let E be a finite equivalence relation. Then for any Polish group G, B 1 (E, G) = Z 1 (E, G). Proof. Let T be a Borel transversal for E. Let α, β ∈ Z 1 (E, G) in order to show that α ∼ β. Define f : X → G as follows: For a given x ∈ X, let {x0 } = [x]E ∩ T . Then put f (x) = β(x0 , x)α(x, x0 ). (Thus f (x) = 1, if x ∈ T .) Then if xEy and {x0 } = T ∩ [x]E = T ∩ [y]E , we have β(x, y) = β(x0 , y)β(x, x0 ) = β(x0 , y)β(x0 , x)−1 = f (y)α(x0 , y)(f (x)α(x0 , x))−1 = f (y)α(x0 , y)α(x, x0 )f (x)−1 = f (y)α(x, y)f (x)−1 . 2 S
Proposition 26.2. Let E1 ⊆ E2 ⊆ . . . , E = n En , G a Polish group. If every cohomology class is dense in each Z 1 (En , G), the same is true for Z 1 (E, G). Similarly, if every B 1 (En , G) is dense in Z 1 (En , G), the same is true for B 1 (E, G) and Z 1 (E, G).
154
III. COCYCLES AND COHOMOLOGY
Proof. Let Γ1 ≤ Γ2 ≤ . . . be a sequence of countable subgroups of (n) (n) Aut(X, µ) such that Γn generates En . Put Γn = {γi }∞ be i=1 . Let µi (n) (n) the copy of µ on graph(γi ) via the R map x 7→ (x, γi (x)). Let M be the usual measure on E, M (A) = card(Ax )dµ(x), and Mn the similar P −i (n) measure for En , so that Mn = M |En . Note that M |En ∼ ∞ i=1 2 µi , M ∼ P∞ P∞ −(i+m) µ(m) . i=1 m=1 2 i Let now α, β ∈ Z 1 (E, G). Consider α|En , β|En . Since ∞ XX
(m)
2−(i+m) µi
∼ M |En ,
m≤n i=1
there is fn : X → G such that ∞ XX
2
−(i+m)
Z
(m)
d(fn · α, β)dµi
≤
m≤n i=1
1 , n
where d ≤ 1 is a compatible metric for G. Therefore ! Z Z ∞ X X X (m) (m) −(i+m) −m −i d(fn · α, β)dµi ≤ d(fn · α, β)dµi 2 2 2 i,m
i=1
m≤n
+
X m>n
2−m ≤
1 + 2−m → 0 n m>n X
as n → ∞. So fn · α → β in Z 1 (E, G). For the case of B 1 (E, G), Z 1 (E, G) apply the preceding argument to α = 1. 2 Corollary 26.3 (Parthasarathy-Schmidt [PS]). If E is hyperfinite and G is Polish, then every cohomology class is dense in Z 1 (E, G). Moreover we have the following result, certain cases of which were proved in Hjorth [Hj1]. Theorem 26.4. Let E be ergodic, hyperfinite, and let G 6= {1} be a Polish group. Then B 1 (E, G) 6= Z 1 (E, G). If moreover G admits an invariant metric, then every cohomology class is dense and meager in Z 1 (E, G) and so ROUGH(E, G) = Z 1 (E, G) = B 1 (E, G). Proof. For the first assertion, assume first that G is not compact. Then by Solecki [So] there is a free continuous action of G on a Polish space Y Y . Let α : E → G be the and a Borel reduction p : X → Y of E to EG corresponding cocycle, α(x, y) · p(x) = p(y). Then α is not in B 1 (E, G), since if α(x, y) = f (y)f (x)−1 , then q(x) = f (x)−1 · p(x) is E-invariant, so q(x) = y0 ∈ Y , a.e., i.e., p maps into a single orbit of G, a.e., a contradiction. When G 6= {1} is compact, one can use for example the fact that there is a cocycle α : E → G with ER(α) = G (see Zimmer [Zi1]). Then it is easy to see that α 6∈ B 1 (E, G).
26. COHOMOLOGY II: THE HYPERFINITE CASE
155
For the second assertion, we have, by Section 23, that B 1 (E, G) is in ROUGH(E, G) and so all cohomology classes in B 1 (E, G) = Z 1 (E, G) are dense and meager. 2 In the next section, we will see more general facts that also imply that the action of L(X, µ, G) on Z 1 (E, G) is turbulent, when G admits an invariant metric (see 27.4). We will next prove a converse result. Theorem 26.5 (Kechris). Let E be given by a free action of a countable group. Then the following are equivalent: (i) E is hyperfinite, (ii) For every countable G, B 1 (E, G) = Z 1 (E, G), (iii) For every Polish G, B 1 (E, G) = Z 1 (E, G). Proof. We have seen that (i) ⇒ (iii) ⇒ (ii). We will now prove that (ii) ⇒ (i). Lemma 26.6. Let a ∈ A(Γ, X, µ) and let ϕ : Γ → G be a homomorphism, where G is a countable group. Then if ϕ ∈ B 1 (a, G), ϕ(Γ) is amenable. Proof. Since ϕ ∈ B 1 (a, G), there is a sequence fn ∈ L(X, µ, G) such that ∀γ∀∞ n(fn (γ · x) = ϕ(γ)fn (x)), a.e.(x). Then define Φ : `∞ (G) → C by Z Φ(F ) = ΦN ({F (fn (x))}n )dµ(x), where ΦN : `∞ (N) → C is a shift-invariant mean (i.e., satisfies ΦN ({xn }) = ΦN ({xn+1 }) which is universally measurable, so that the above integral makes sense. We are using here the theorem of Mokobodzki that asserts the existence of such means under the Continuum Hypothesis (CH) (see, e.g., Kechris [Kec1]). Since the statement of the result we want to prove is sufficiently absolute, it is well-known that we can eliminate the use of CH by the usual metamathematical methods. See, e.g., Adams-Lyons [AL] for an exposition of this metamathematical technique. It is not hard to see now that Φ is ϕ(Γ)-invariant under the translation action of ϕ(Γ) on G, thus ϕ(Γ) is amenable. 2 So let E = EΓX , where Γ is a countable group acting freely on X. Call this action a. Let ϕ : Γ → Γ be the identity. Then Z 1 (a, Γ) can be identified with Z 1 (E, Γ) and similarly for B 1 (a, Γ) and B 1 (E, Γ), and we have that 2 ϕ ∈ B 1 (a, Γ), so ϕ(Γ) = Γ is amenable and thus E is hyperfinite. I do not know if the hypothesis that E is given by a free action is needed in 26.5. Before we proceed to the next result, it would be convenient to state a characterization of the cocycles in B 1 (E, G) for an arbitrary equivalence relation E and countable group G.
156
III. COCYCLES AND COHOMOLOGY
For each countable group G, consider the equivalence relation on GN given by {gn }E0 (G){hn } ⇔ ∀∞ n(gn = hn ). Clearly E0 (G) is Borel hyperfinite. Let X0 (G) = GN /E0 (G). The group G acts freely on X0 (G) by diagonal (left-) translation g · [{gn }] = [{ggn }]. X (G)
The equivalence relation EG 0 alence relation
, when lifted up to GN , becomes the equiv-
{gn }Ft,0 (G){hn } ⇔ ∃g∀∞ n(ggn = hn ). S Note that Ft,0 (G) is hyperfinite. Indeed Ft,0 (G) = ∞ m=1 Fm , where {gn }Fm {hn } ⇔ ∃g∀n ≥ m(ggn = hn ). Let G act (diagonally) by translation on GN g · {gn } = {ggn }. Denote by Ft (G) the corresponding equivalence relation. It is clearly smooth as {1} × GN is a Borel transversal. Now Fm is Borel reducible to Ft (G), so it is smooth as well. Since F1 ⊆ F2 ⊆ . . . , Ft,0 (G) is hyperfinite (see Dougherty-Jackson-Kechris [DJK], 5.1). Given now an equivalence relation E on (X, µ) and a homomorphism X (G) p : X → X0 (G) of E into EG 0 , a.e., with Borel lifting, we have an associated cocycle α ∈ Z 1 (E, G) given by α(x, y)·p(x) = p(y). Equivalently, by lifting, for any Borel homomorphism q of E into Ft,0 (G), a.e., we have an associated cocycle α ∈ Z 1 (E, G) defined by α(x, y) = the unique g ∈ G such that ∀∞ n(gq(x)n = q(y)n ), where q(x) = {q(x)n }. We now have: Proposition 26.7. Let E be an equivalence relation on (X, µ) and G a countable group. Then α ∈ B 1 (E, G) iff α is associated to an a.e. Borel homomorphism of E into Ft,0 (G). Proof. ⇒: If α ∈ B 1 (E, G), we can find Borel qn : X → G such that, a.e., xEy ⇒ ∀∞ n(qn (y) = α(x, y)qn (x)). Let q(x) = {qn (x)}, q : X → GN . Then clearly q is a Borel homomorphism of E into Ft,0 (G) with associated cocycle α. ⇐: Let q : X → GN be a Borel homomorphism of E into Ft,0 (G) with associated cocycle α and let q = {qn }. Then, a.e., xEy ⇒ ∀∞ n(qn (y) = α(x, y)qn (x)), so α ∈ B 1 (E, G). 2 In the case of actions a ∈ A(Γ, X, µ), we say that α ∈ Z 1 (a, G) is associated to a Borel homomorphism of a into Ft,0 (G) if there is a Borel map q : X → GN such that, a.e., q(γ · x)Ft,0 (G)q(x) and α(γ, x) = the
26. COHOMOLOGY II: THE HYPERFINITE CASE
157
unique g ∈ G such that ∀∞ n(gq(x)n = q(γ · x)n ). Then the analog of 26.7 goes through. We now have the following consequence. Proposition 26.8 (Kechris). Let E be E0 -ergodic and G be a countable group. Then B 1 (E, G) is closed. Similarly for actions. Proof. Let α ∈ B 1 (E, G). Then α is associated to a homomorphism q of E into Ft,0 (G). Since E is E0 -ergodic and Ft,0 (G) is hyperfinite, q = {qn } maps into a single Ft,0 (G)−class, a.e., say that of {h0n }. Thus ∃g∀∞ n(qn (x) = gh0n ), a.e.(x), so there is g : X → G such that ∀∞ n(qn (x) = g(x)h0n ), a.e.(x). Thus we have, on a set of measure 1, xEy ⇒ ∀∞ n(qn (y) = α(x, y)qn (x) = g(y)h0n = α(x, y)g(x)h0n ), therefore xEy ⇒ α(x, y) = g(y)g(x)−1 , i.e., α ∈ B 1 (E, G) and B 1 (E, G) is closed.
2
Schmidt has proved 26.8 when G is an abelian locally compact Polish group; see Schmidt [Sc5], 3.3. We now have the following result that generalizes a theorem of Schmidt (see Schmidt [Sc3], Remark (3.5)), who dealt with the case of abelian G. Theorem 26.9 (Kechris). Let Γ be a countable group. Then the following are equivalent: (i) Γ is amenable, (ii) ∀a ∈ FR(Γ, X, µ)∀Polish G, B 1 (a, G) = Z 1 (a, G), (iii) ∀a ∈ FR(Γ, X, µ)∀countable G, B 1 (a, G) = Z 1 (a, G), (iv) ∀a ∈ FRERG(Γ, X, µ)∀Polish G 6= {1}, B 1 (a, G) 6= B 1 (a, G) = Z 1 (a, G), (v) ∀a ∈ FRERG(Γ, X, µ)∃countable G, B 1 (a, G) 6= B 1 (a, G). Proof. (i) ⇒ (ii) by 26.3. (i) ⇒ (iv) by 26.4. (ii) ⇒ (iii) and (iv) ⇒ (v) are obvious. Finally, we prove ¬(i) ⇒ ¬(iii), ¬(v).
158
III. COCYCLES AND COHOMOLOGY
Assume that Γ is not amenable. Then the shift action s of Γ on 2Γ is free and E0 -ergodic (see, e.g. Hjorth-Kechris [HK3], A4.1). Then B 1 (s, G) is closed by 26.8, for any countable G, and B 1 (s, Γ) 6= Z 1 (s, Γ) by 26.6. 2 Comments. There is an extensive literature on the structure of cocycles of hyperfinite ergodic equivalence relations E, especially in the works of Bezuglyi, Danilenko, Fedorov, Golodets, Schmidt, Sinelshchikov and Zimmer. For example, it is known that the Polish locally compact groups G that are essential ranges of such cocycles are exactly the amenable ones. Given such a G, the set of cocycles α ∈ Z 1 (E, G) such that ER(α) = G is comeager in Z 1 (E, G). Every ergodic, quasi-invariant action of an amenable locally compact Polish group is isomorphic to a Mackey action (in the sense of Section 21, (C)). Every cocycle α ∈ Z 1 (E, G), G Polish, is cohomologous to a cocycle β taking values in any given countable dense subgroup of G. Finally, extending Dye’s Theorem, for any Polish group G, if α, β ∈ Z 1 (E, G) have essential range G, then α, β are weakly equivalent, α ∼w β. Since we are primarily interested here in non-hyperfinite equivalence relations, we will not pursue the hyperfinite case any further. For a recent survey of results in this case, see Bezuglyi [Be]. 27. Cohomology III: The non E0 -ergodic case (A) We start with the following result: Proposition 27.1 (Kechris). Let E be ergodic but not E0 -ergodic and let G be a Polish group admitting an invariant metric. If α ∈ Z 1 (E, G) has ˜ on Z 1 (E, G), i.e., G ˜ α 6= {1}, then non-trivial stabilizer for the action of G α ∈ ROUGH(E, G). Similarly for cocycles of actions. ˜ α , g0 6= 1. Recall also from the proof of 25.1 that for Proof. Fix g0 ∈ G ˜ g), a.e., where d ≤ 1 is an invariant ˜ f, g ∈ Gα , we have d(f (x), g(x)) = d(f, metric on G. Fix also a countable group Γ and a Borel action of Γ on X with E = EΓX . Since E is not E0 -ergodic, there is a sequence {An } of Borel sets with ˜ be defined by µ(γ · An 4An ) → 0, ∀γ ∈ Γ, and µ(An ) = 1/2. Let χn ∈ G χn (x) = g0 (x), if x ∈ An , χn (x) = 1, if x 6∈ An . Claim. χn · α → α. Granting this claim, if α ∈ SMOOTH(E, G), towards a contradiction, ˜ = [α]∼ is closed, so by Effros’ Theorem, writing H = G ˜ α, then the orbit G·α ˜ ˜ we R have that χn H → H in G/H, thus we can find hn ∈ H with d(χn , hn ) = d(χn (x), hn (x))dµ(x) → 0. Clearly d(χn (x), hn (x)) = d(g0 (x), hn (x)), if x ∈ An ; = d(1, hn (x)), if x 6∈ An . But 1, g0 , hn ∈ H, so ˜ 0 , hn ) = d(g0 (x), hn (x)), d(g ˜ hn ) = d(1, hn (x)), d(1,
27. COHOMOLOGY III: THE NON E0 -ERGODIC CASE
159
a.e., thus ( ˜ 0 , hn ), if x ∈ An , d(g d(χn (x), hn (x)) = ˜ d(1, hn ), if x ∈ / An , a.e., and so ˜ n , hn ) = 1 (d(g ˜ 0 , hn ) + d(1, ˜ hn )) → 0, d(χ 2 i.e., hn → g0 , hn → 1, a contradiction. Proof of the Claim. We need to show that for each γ, Z (*) d(χn (γ −1 · x)α(x, γ −1 · x)χn (x)−1 , α(x, γ −1 · x))dµ(x) converges to 0 as n → ∞. But ( g0 (γ −1 · x), if x ∈ γ · An , χn (γ −1 · x) = 1, if x 6∈ γ · An , so d(χn (γ −1 ·x)α(x, γ −1 ·x)χn (x)−1 , α(x, γ −1 ·x)) ≤ 1, if x ∈ γ ·An 4An , and otherwise it is equal to d(g0 (γ −1 · x)α(x, γ −1 · x)g0 (x)−1 , α(x, γ −1 · x)) = 0, if x ∈ γ · An ∩ An (as g0 · α = α) and to d(α(x, γ −1 · x), α(x, γ −1 · x)) = 0, if x 6∈ γ · An ∪ An . Thus (*) ≤ µ(γ · An 4An ) → 0. 2 Corollary 27.2. For any E which is ergodic but not E0 -ergodic and any Polish group G 6= {1} admitting an invariant metric, B 1 (E, G) ⊆ ROUGH(E, G). Similarly for actions. ˜ 1 = G 6= {1}. Proof. G
2
Thus for any ergodic, non E0 -ergodic E and non-trivial G admitting an invariant metric (in particular countable G), the rough part ROUGH(E, G) of Z 1 (E, G) is non-empty. Of course SMOOTH(E, G) can be empty and ROUGH(G) can contain a single cohomology class closure, as for example when E is hyperfinite. We will discuss later in this section various examples where ROUGH(E, G) contains many cohomology class closures and similarly for SMOOTH(E, G). ˜ on each cohomology class We have seen in Section 23, that the action of G closure C contained in ROUGH(E, G) is minimal and every cohomology class in C is meager in C. We actually have the following result. Theorem 27.3 (Kechris). Suppose E is not E0 -ergodic and let G be a Polish group admitting an invariant metric. Let α ∈ ROUGH(E, G). Then ˜ on the closure of [α]∼ is turbulent. Similarly for actions. the action of G Proof. The argument is reminiscent of that in the proofs of 5.2 and 13.3. We only need to verify that α is a turbulent point for this action. Fix
160
III. COCYCLES AND COHOMOLOGY
X a countable group Γ = {γk }∞ k=1 and a Borel action of Γ on X with EΓ = E. Consider the open nbhd Z X 1 −k ˜ Uα, = {β : d (α, β) = 2 d(α(x, γk · x), β(x, γk · x))dµ(x) < }, k
˜ where > 0 and d ≤ 1 is a compatible invariant metric for G. Let f ∈ G be such that β = f · α ∈ Uα, . It is enough to find a continuous path ˜ λ 7→ fλ , λ ∈ [0, S1], from 1 to f inN G with fλ · α lying in Uα, , ∀λ. Let E0 = n≥1 En on Y = 2 , where each En is a finite Borel equivalence relation and E1 ⊆ E2 ⊆ . . . Since E is not E0 -ergodic, let π : X → Y be a Borel homomorphism of E into E0 such that µ(π −1 (C)) = 0 for every E0 -class C. P Say d˜1 (α, β) = ρ < and choose N0 large enough so that k>N0 2−k < (−ρ) 3 .
Now for almost all x, ∃n ≥ 1∀k ≤ N0 (π(x)En π(γk · x)En π(γk−1 · x)).
So if Xn = {x : ∀k ≤ N0 (π(x)En π(γk · x))}, then X1 ⊆ X2 ⊆ . . . and µ(Xn ) → 1. So choose N ≥ N0 large enough so that ( − ρ) . µ(∼ XN ) < 3 Then as in the proof of 5.2 we can find a continuous map λ 7→ Xλ from [0,1] to MALGµ with X0 = ∅, X1 = X, λ ≤ λ0 ⇒ Xλ ⊆ Xλ0 & µ(Xλ0 \ Xλ ) ≤ λ0 − λ, and Xλ π −1 (EN )-invariant, so that, in particular, if k ≤ N0 and x ∈ XN , then x ∈ Xλ ⇔ γk · x ∈ Xλ . Let fλ (x) = f (x), if x ∈ Xλ ; = 1, if x 6∈ Xλ . We claim that fλ · α ∈ Uα, . Since f0 = 1, f1 = f and λ 7→ fλ is continuous, this will complete the proof. Now d˜1 (α, fλ · α) is clearly bounded by Z X −k 2 d(α(x, γk · x), fλ (γk · x)α(x, γk · x)fλ (x)−1 )dµ(x) k≤N0
with an error at most (−ρ) 3 . We can break this sum into four parts, restricting the integral to the sets Xλ ∩ XN , Xλ \ XN , XN \ Xλ , ∼ Xλ ∩ ∼ XN , resp. The first part is bounded by d˜1 (α, β) = ρ, since fλ = f on Xλ . The second is bounded by µ(∼ XN ) < (−ρ) 3 , since d ≤ 1. The third is 0 since fλ = 1 on ∼ Xλ . Finally the last one is again bounded by µ(∼ XN ). Thus d˜1 (α, fλ · α) is less than (−ρ) + ρ + (−ρ) + 0 + (−ρ) = . 2 3 3 3 Remark. This argument actually shows that for any Polish group G and any E which is not E0 -ergodic, every α ∈ Z 1 (E, G) is a turbulent point ˜ on the closure of the cohomology class of α. for the action of G Corollary 27.4. Suppose E is ergodic but not E0 -ergodic and let G 6= {1} ˜ on be a Polish group admitting an invariant metric. Then the action of G
27. COHOMOLOGY III: THE NON E0 -ERGODIC CASE
161
B 1 (E, G) is turbulent. In particular, the cohomology relation on B 1 (E, G) (and thus also on Z 1 (E, G)) does not admit classification by countable structures. Similarly for actions. We should point out here that Hjorth [Hj1] has proved certain cases of 27.4 by different methods for hyperfinite E. (B) We will now discuss some implications of the results in (A) to cocycle superrigidity. Let a ∈ A(Γ, X, µ) and let G be a class of groups. Then a is called G-cocycle superrigid if every α ∈ Z 1 (a, G), where G ∈ G, is cohomologous to a homomorphism ϕ : Γ → G. This concept arose in recent work of Popa, where many important examples of actions that are Gcount -cocycle superrigid were discovered, where Gcount = the class of all countable groups. For instance, Popa [Po2] shows (among other things) that the shift action of Γ on [0, 1]Γ , for various kinds of groups Γ, is Gcount -superrigid. When G = {G} we say that a is G-cocycle superrigid. We now have the following result which shows that superrigidity entails E0 -ergodicity. Corollary 27.5. Let a ∈ WMIX(Γ, X, µ) (in fact it is enough to assume that every finite index subgroup Γ acts ergodically) and a 6∈ E0 RG(Γ, X, µ). Then for any countable group G 6= {1}, a is not G-cocycle superrigid. In fact, for every α ∈ ROUGH(a, G), the generic cocycle in the closure of the cohomology class of α is not cohomologous to a homomorphism. ˜ on C Proof. Let α ∈ ROUGH(a, G), C = [α]∼ . Then the action of G is turbulent. Let A = {β ∈ C : β is cohomologous to a homomorphism}. ˜ on C is minimal, A is either meager or comeager in Since the action of G C. Assume towards a contradiction that A is comeager. Then by going if necessary to comeager subset of A, we can assume that there is a Borel function Φ : A → Hom(Γ, G) (viewed as a Polish space with the product topology) such that α ∼ Φ(α). But then, by 20.1, α ∼ β ⇔ Φ(α) ∼ Φ(β) ⇔ Φ(α), Φ(β) are conjugate. Clearly conjugacy in Hom(Γ, G) is induced by a countable group action, so ∼ on A is classifiable by countable structures, contradicting turbulence. 2 We also have the following consequence concerning generic actions. Corollary 27.6. Suppose the group Γ does not have property (T). Then the generic action of Γ is not G-cocycle superrigid, for any countable group G 6= {1}. Proof. The generic action of Γ is weak mixing and non E0 -ergodic (by 12.3 and 12.9). 2 We will next see some characterizations of the class of property (T) groups, related to superrigidity. Hjorth-Kechris [HK2], B.4 and (independently) Jolissaint [Jo2] proved that if Γ has property (T), and G is a Polish locally compact group with
162
III. COCYCLES AND COHOMOLOGY
HAP, then for every a ∈ ERG(Γ, X, µ), every cocycle α ∈ Z 1 (a, G) is cohomologous to a cocycle β which takes values in a compact subgroup of G. (See also Connes-Jones [CJ2] for earlier related results in the context of operator algebras.) In particular, if G contains no non-trivial such subgroups (e.g., G countable, torsion free), then α ∈ B 1 (a, G), so a is also G-cocycle superrigid. We note the following converse. Corollary 27.7. Let Γ be a countable group. Then the following are equivalent: (i) Γ does not have property (T). (ii) For every countable group G 6= {1}, the generic ergodic action of Γ is not G-cocycle superrigid. (iii) There is a countable, torsion-free group G with HAP and an ergodic action of Γ which is not G-cocycle superrigid. Proof. (i) ⇒ (ii) by 27.6. (ii) ⇒ (iii) is obvious. Finally ¬(i) ⇒ ¬(iii) by the preceding paragraph. 2 Call now a countable group Γ somewhere cocycle superrigid if every ergodic action of Γ is G-superrigid for some countable G 6= {1}, perhaps 0 depending on the action. Denote by GHAP the class of Polish locally compact groups that have HAP and contain no non-trivial compact subgroups. Finally we say that Γ is G-cocycle superrigid if every ergodic action of Γ is G-cocycle superrigid. Corollary 27.8. Let F be a class of countable groups. Then the following are equivalent: (i) F is contained in the class of property (T) groups, (ii) ∀Γ ∈ F(Γ is somewhere cocycle superrigid), 0 (iii) ∀Γ ∈ F(Γ is GHAP -cocycle superrigid). Proof. (i) ⇒ (iii) has been already discussed and (iii) ⇒ (ii) is obvious. Assume now (i) fails and let Γ ∈ F fail to have property (T). Then by 27.6, the generic action a of Γ is not G-cocycle superrigid, for any countable G 6= {1}, so (ii) fails. 2 Thus the class of property (T) groups is the largest class of countable groups all of whose ergodic actions have some non-trivial cocycle superrigidity. (C) As in Section 15, the turbulence results 27.3, 27.4 above have also consequences concerning path connectedness in the space of cocycles. Theorem 27.9 (Kechris). Suppose E is not E0 -ergodic and let G be a Polish group admitting an invariant metric. Then for any α ∈ Z 1 (E, G) the closure of the cohomology class of α is path connected. Similarly for actions. ˜ · α and G ˜ is Proof. If α ∈ SMOOTH(E, G), this is clear as [α]∼ = G path connected. If α ∈ ROUGH(E, G), then this follows from the proof of 27.3 and the proof of 15.2. 2
27. COHOMOLOGY III: THE NON E0 -ERGODIC CASE
163
We also have the following. Theorem 27.10 (Kechris). For any equivalence relation E on (X, µ) and countable G, B 1 (E, G) is path connected. Similarly for actions. Proof. If E is not E0 -ergodic, this follows from 27.9. If E is E0 -ergodic, ˜ · 1, and we are done. 2 then by 26.8, B 1 (E, G) = B 1 (E, G) = G Corollary 27.11. For any Γ and countable group G, {(a, α) ∈ AZ 1 (Γ, X, µ, G) : α ∈ B 1 (a, G)} is path connected. Proof. Fix (a, α) in this set. Then α is path connected to 1 with a path in B 1 (a, G). So it is enough to show that A(Γ, X, µ)×{1} is path connected, which follows from 15.11. 2 Corollary 27.12. If Γ is amenable, then AZ 1 (Γ, X, µ, G) is path connected, for any countable group G. (D) We have already seen that when E is ergodic but not E0 -ergodic and G 6= {1} is countable, ROUGH(E, G) is always non-empty, while if E is moreover hyperfinite, then ROUGH(E, G) = Z 1 (E, G). We also have: Proposition 27.13. Let Γ be a countable group. Then for any action a ∈ ERG(Γ, X, µ) \ E0 RG(Γ, X, µ) and any weakly commutative, nontrivial, countable group G, ROUGH(a, G) = Z 1 (a, G). Proof. Let α ∈ Z 1 (a, G), where a ∈ ERG(Γ, X, µ) \ E0 RG(Γ, X, µ). We want to show that α ∈ ROUGH(a, G). By 27.1, we can assume that ˜ α is trivial. Thus by Effros’ Theorem it is enough to find the stabilizer G ˜ gn ∈ G such that gn · α → α but gn 6→ 1. Let {gn } ⊆ G witness the weak commutativity of G, i.e., gn 6= 1 and ∀g ∈ G∀∞ n(ggn = gn g). Then clearly ˜ gn · α → α. On the other hand gn 6→ 1 in G. 2 This should be contrasted with the following result. Proposition 27.14 (Kechris). Assume that Γ is a group that is not inner amenable and let a ∈ ERG(Γ, X, µ). Let α0 (γ, x) = γ, so that α0 ∈ Z 1 (a, Γ). Then α0 ∈ SMOOTH(a, Γ), thus ROUGH(a, Γ) 6= Z 1 (a, Γ). ˜ α = {1}. Indeed, if f · α0 = α0 , then f (γ · x) = Proof. First note that Γ 0 −1 γf (x)γ , so f : X → Γ is a homomorphism of the Γ-action on X to the conjugacy action on Γ. Thus f∗ µ is an ergodic, invariant measure for this last action, so, since Γ is not inner amenable, it concentrates on {1}, i.e., f = 1. ˜ Thus, by Effros’ Theorem and 23.5, it is enough to show that if fn ∈ Γ and fn · α0 → α0 , then fn → 1. Equivalently, it is enough to show that if ∀γ∀∞ n(fn (γ · x)γfn (x)−1 = γ), a.e.(x), then fn → 1. So assume that (*)
∀γ∀∞ n(fn (γ · x) = γfn (x)γ −1 ), a.e.(x),
164
III. COCYCLES AND COHOMOLOGY
but fn 6→ 1, towards a contradiction. Then the set A = {x : ∃∞ n(fn (x) 6= 1)} has positive measure, but it is also Γ-invariant by (*), thus, by ergodicity, (**)
∃∞ n(fn (x) 6= 1), a.e.(x).
Put ϕ(x) = {fn (x) : n ∈ N}. Claim. ϕ(x) is finite, a.e.(x). We assume this claim and complete the proof. Let Ψ(x) = {g ∈ ϕ(x) : g 6= 1 & ∃∞ n(fn (x) = g)}. Thus by (∗∗ ), 1 6∈ Ψ(x) 6= ∅. Next we check that Ψ(γ · x) = γΨ(x)γ −1 . Indeed fix γ, x and by (*) let N be such that ∀n ≥ N (fn (γ ·x) = γfn (x)γ −1 ). If g ∈ Ψ(x), let N < n0 < n1 < . . . be such that fni (x) = g. Then γgγ −1 = γfni (x)γ −1 = fni (γ · x), so γgγ −1 ∈ Ψ(γ · x). Thus in particular x 7→ card(Ψ(x)) is Γ-invariant, so constant a.e., say equal to K ≥ 1. Then Ψ(x) is a homomorphism of the Γ-action on X into the conjugacy action of Γ on the set PK (Γ) of subsets of Γ of cardinality K. Thus Ψ∗ µ is an invariant, ergodic measure for this action, so concentrates on a single finite orbit, thus {γΨ(x)γ −1 : γ ∈ Γ} is finite, a.e.(x). −1 : γ ∈ Γ} is infinite (as Γ is Now fix x and gS 0 ∈ Ψ(x). Then C0 = {γg0 γ −1 ICC), but C0 ⊆ γ∈Γ γΨ(x)γ which is finite, a contradiction.
Proof of the Claim. Let E(x) be the equivalence relation on N given by mE(x)n ⇔ fm (x) = fn (x). Then, by (*), ∀γ∃N ∀m, n ≥ N [mE(x)n ⇔ mE(γ · x)n], a.e.(x). In particular, ϕ(x) is finite iff ϕ(γ · x) is finite, so, towards a contradiction, we can assume that ϕ(x) is infinite, a.e.(x). Let then n1 (x) < n2 (x) < . . . enumerate in increasing order the least elements of the E(x)-classes at which fn (x) 6= 1. Again we must have ∀γ∀∞ i(ni (x) = ni (γ · x)). Now define a mean Φ : `∞ (Γ) → C by Z Φ(F ) = ΦN (i 7→ F (fni (x) (x)))dµ(x), where ΦN : `∞ (N) → C is a shift invariant mean as in the proof of 26.6. We will verify that Φ is conjugacy invariant. Denote by γ · F the conjugacy R action of Γ on `∞ (Γ) : γ · F (δ) = F (γ −1 δγ). So Φ(γ −1 · F ) = ΦN (i 7→ F (γfni (x) (x)γ −1 ))dµ(x). For each γ, x, if i is sufficiently large, γfni (x) (x)γ −1 = fni (x) (γ · x) = fni (γ·x) (γ · x), thus ΦN (i 7→ F (γfni (x) (x)γ −1 )) = ΦN (i 7→ F (fni (γ·x) (γ · x))),
27. COHOMOLOGY III: THE NON E0 -ERGODIC CASE
165
and so Φ(γ
−1
Z · F) =
ΦN (i 7→ F (fni (γ·x) (γ · x)))dµ(x) Z
=
ΦN (i 7→ F (fni (x) (x)))dµ(x)
= Φ(F ). Since Γ is not inner amenable, Φ must concentrate on {1}, i.e., Φ(χΓ\{1} ) = 0, so Z ΦN (i 7→ χΓ\{1} (fni (x) (x)))dµ(x) = 0. But for each i, x, fni (x) (x) 6= 1, so Z ΦN (i 7→ χΓ\{1} (fni (x) (x)))dµ(x) = 1, which is a contradiction.
2
Remark. The previous proof also shows that if θ : Γ → G is a homomorphism from Γ to a countable group G and G is not inner amenable relative to θ(Γ) (i.e., every finitely additive probability measure on G, which is invariant under conjugation by elements of θ(Γ), concentrates on 1), then θ ∈ SMOOTH(a, G), for any a ∈ ERG(Γ, X, µ). Corollary 27.15. Let Γ, G be countable groups, a ∈ FRERG(Γ, X, µ), b ∈ FRERG(G, Y, ν) and let απ ∈ Z 1 (a, G) be an orbit equivalence cocycle arising from an orbit equivalence π : X → Y of a and b (i.e., απ (γ, x) · π(x) = π(γ · x)). If G is not inner amenable, then απ ∈ SMOOTH(a, G). Proof. We can identify Z 1 (a, G) with Z 1 (Ea , G) and Z 1 (b, G) with b , G). Since π is an isomorphism of Ea with Eb it induces a homeo˜ morphism π ∗ between Z 1 (Ea , G) and Z 1 (Eb , G), preserving the G-actions, defined by π ∗ (α)(x, y) = α(π −1 (x), π −1 (y)). Then π ∗ (απ ) = α0 , where α0 ∈ Z 1 (b, G) is given by α0 (g, x) = g. Since α0 ∈ SMOOTH(b, G), clearly απ ∈ SMOOTH(a, G). 2 Z 1 (E
(E) We will next discuss some other examples of behavior of the dichotomy ROUGH/SMOOTH. First let G be a Polish group, E, F be equivalence relations on the spaces (X, µ), (Y, ν), resp., and consider the product E × F on (X × Y, µ × ν). Then we have a canonical lift map α ∈ Z 1 (E, G) 7→ α ˆ ∈ Z 1 (E × F, G), given by α ˆ ((x1 , y1 )(x2 , y2 )) = α(x1 , x2 ). Clearly it is continuous. Also if gn ∈ L(X, µ, G) is such that gn · α → β, then, letting gˆn ∈ L(X × Y, µ × ν, G) ˆ Conversely, be defined by gˆn (x, y) = gn (x), we obviously have gˆn · α ˆ → β. if fn ∈ L(X × Y, µ × ν, G) and fn · α ˆ → βˆ pointwise a.e., then, letting y y fn (x) = fn (x, y), we have fn · α → β pointwise a.e., ν-a.e.(y).
166
III. COCYCLES AND COHOMOLOGY
Indeed, dropping null sets, we can assume that α, β are strict cocycles and fn is defined everywhere. Let then A ⊆ X × Y be Borel (E × F )invariant with µ × ν(A) = 1, and such that for (xi , yi ) ∈ A, x1 Ex2 , y1 F y2 , we have fn (x2 , y2 )α(x1 , x2 )fn (x1 , y1 )−1 → β(x1 , x2 ). Then, by Fubini, for almost all y we have that (x, y) ∈ A, a.e.(x). For such y we will show that fny · α → β pointwise, a.e. Indeed, fix a Borel E-invariant set B ⊆ X with µ(B) = 1 such that ∀x ∈ B(x, y) ∈ A. Then if x1 Ex2 , x1 , x2 ∈ B, we have (x1 , y), (x2 , y) ∈ A and (x1 , y)E × F (x2 , y), so fn (x2 , y)α(x1 , x2 )fn (x1 , y)−1 → β(x1 , x2 ), i.e., fny · α → β pointwise on B. It follows that: ˆ (i) α ∼ β ⇔ α ˆ ∼ β, ˆ ∼, (ii) α ∈ [β]∼ ⇔ α ˆ ∈ [β] (iii) L(X, µ, G)α = {1} ⇔ L(X × Y, µ × ν, G)αˆ = {1}, (iv) L(X, µ, G)α = {1} & α ∈ SMOOTH(E, G) ⇔ L(X × Y, µ × ν, G)αˆ = {1} & α ˆ ∈ SMOOTH(E × F, G) (For (iv), ⇐ follows from (i), (iii) and the continuity of α 7→ α ˆ ; ⇒ follows from (iii), Effros’ Theorem and Fubini.) The following fact will be a consequence of results in Section 29: There is an ergodic equivalence relation E on (X, µ) and a countable group G such that SMOOTH(E, G) = Z 1 (E, G), and there are 2ℵ0 many α ∈ Z 1 (E, G) pairwise not cohomologous such that L(X, µ, G)α = {1} and also 2ℵ0 many β ∈ Z 1 (E, G) pairwise not cohomologous such that L(X, µ, G)β 6= {1}. (The equivalence relation E is induced by the shift action of the free group with ℵ0 generators F∞ on 2F∞ . It will be shown in 29.8 that SMOOTH(E, F∞ ) = Z 1 (E, F∞ ). Now there are 2ℵ0 many pairwise non-conjugate homomorphisms ϕ : F∞ → F∞ with ϕ(F∞ ) = F∞ and thus by 25.4, (ii) the stabilizer of ϕ ∈ Z 1 (E, F∞ ) is equal to {1}. Also there are 2ℵ0 many pairwise non-conjugate homomorphisms ψ : F∞ → F∞ and h 6= 1 in F∞ such that ψ(F∞ ) ⊆ {hn : n ∈ Z}. Then h ∈ L(X, µ, F∞ )ψ 6= {1}.) ˆ = E × F , F ergodic, hyperfinite, which is clearly not E0 Let now E ergodic. Then, by (iii) above and 27.1, for every β ∈ Z 1 (E, G) for which ˆ G) and if β1 6∼ β2 , then by (ii) above L(X, µ, G)β 6= {1}, βˆ ∈ ROUGH(E, [βˆ1 ]∼ 6= [βˆ2 ]∼ , while if α ∈ Z 1 (E, G) is such that L(X, µ, G)α = {1}, then ˆ G). It follows that for E ˆ there are (iv) above shows that α ˆ ∈ SMOOTH(E, ℵ ˆ G) and 2ℵ0 2 0 many distinct closures of cohomology classes in ROUGH(E, ˆ G). many distinct (of course closed) cohomology classes in SMOOTH(E, (F) We next consider another example in which the rough part contains 2ℵ0 many closures of cohomology classes.
27. COHOMOLOGY III: THE NON E0 -ERGODIC CASE
167
Let a0 , a1 , . . . be free generators of F∞ , and let E = EFX∞ be given by a free, weakly mixing but non E0 -ergodic action of F∞ on (X, µ). Let G = F∞ × Z2 . Let F ⊆ p(N) be a family of 2ℵ0 many almost disjoint infinite subsets of N (i.e., X 6= Y ∈ F ⇒ X ∩ Y is finite). For X ∈ F define the homomorphism ϕX : F∞ → G by ϕX (an ) = (an , 0), if n ∈ X; = (1, 0), if n 6∈ X. First we claim that ϕX ∈ ROUGH(E, G). Indeed, this follows from 25.4, (ii), and 27.1 as CG (ϕX (F∞ )) ⊇ Z2 6= {1}. Finally, we claim that if X 6= Y, ϕX 6∈ [ϕY ]∼ . Assume not, towards a contradiction, and find fn ∈ L(X, µ, G) with fn · ϕY → ϕX pointwise a.e. Thus (*)
∀γ∀∞ n(fn (γ · x)ϕY (γ)fn (x)−1 = ϕX (γ)), a.e.(x).
Define then a mean Φ : `∞ (G) → C by Z Φ(F ) = ΦN ({F (fn (x))})dµ(x), where we use the notation and conventions of the proof of 26.6. If F∞ acts on G by: γ · g = ϕX (γ)gϕY (γ)−1 , and thus on `∞ (G) via γ · F (g) = F (γ −1 · g), it is easy to check, using (*), that Φ is invariant under this action. If ρ(A) = Φ(χA ), is the associated finitely additive probability measure on the power set of G, then ρ is invariant under the map A 7→ ϕX (γ)AϕY (γ)−1 , ∀γ ∈ Γ. If now n ∈ X\Y , then ϕX (an ) = (an , 0), ϕY (an ) = (1, 0), so ϕX (an )AϕY (an )−1 = (an , 0)A. Then if H = han : n ∈ X\Y i ≤ F∞ , so that H × {0} ≤ G, then clearly ρ is H × {0}-invariant under lefttranslation, so H is amenable, a contradiction. (G) We will conclude this section with some open problems. Below we will assume that the equivalence relation E or the action a ∈ A(Γ, X, µ) is ergodic but not E0 -ergodic and G is a non-trivial countable group. We have seen in Section 25 that ROUGH(E, G) and SMOOTH(E, G) are Borel sets. What can be said about their Borel complexity? ˜ α = {1} and G ˜·α We note here that by 27.1, α ∈ SMOOTH(E, G) iff G 1 ˜ α = {1} and ∀∃δ∀n(d˜ (fn · α, α) < δ ⇒ is Gδ iff (by Effros’ Theorem) G ˜ ˜ d is the discrete metric on G and , δ d(fn , 1) < ), where {fn } is dense in G, vary over positive rationals. Therefore SMOOTH(E, G) is the intersection ˜ α = {1}} and a Π0 set. of the Borel set {α ∈ Z 1 (E, G) : G 3 Next we have questions concerning category. Can SMOOTH(E, G) be non-meager? Can there be a cohomology class in SMOOTH(E, G) which is non-meager and thus clopen? Can the set of cocycles in Z 1 (a, G) that are cohomologous to homomorphisms be non-meager? Can the set of reduction cocycles in a given [α]∼ ⊆ ROUGH(E, G) be non-meager in [α]∼ ? Can the set of reduction cocycles be non-meager in Z 1 (E, G)? Concerning the size of ROUGH(E, G), SMOOTH(E, G), let r, resp., s, be the cardinality of the set of cohomology class closures contained in the set ROUGH(E, G), resp., SMOOTH(E, G). Thus always r ≥ 1. We have seen examples, where r = 1, s = 0; r = 2ℵ0 , s = 2ℵ0 . What possible pairs (r, s)
168
III. COCYCLES AND COHOMOLOGY
are realized by appropriately choosing E, G? For instance, is it possible for r, s to be finite or countable and r > 1? We have seen that every cohomology class closure contained in the set ROUGH(E, G) is path connected. How about ROUGH(E, G) and Z 1 (E, G)? Note here that if 1 < r ≤ ℵ0 , then ROUGH(E, G) cannot be path connected, since by Sierpinski’s Theorem (see van Mill [V], Ex. §4.6, 1) one cannot write the interval [0,1] as a non-trivial disjoint union of countably many closed sets. Finally, what can be said about the class of groups Γ without property (T) such that for every a ∈ ERG(Γ, X, µ) \ E0 RG(Γ, X, µ) and countable, non-trivial G, we have Z 1 (a, G) = ROUGH(a, G)? By 26.4 it contains all amenable groups and by 27.14 it is contained in the class of inner amenable groups. 28. The minimal condition on centralizers (A) We discuss here a class of groups that will play an important role in the next section. Let G be a group. For any A ⊆ G we denote by CG (A) = {g ∈ G : ∀a ∈ A(ga = ag)} the centralizer of A in G. A centralizer in G is a subgroup of the form CG (A) for some A ⊆ G, which we can clearly always take to be a subgroup of G. A group G satisfies the minimal condition on centralizers if there is no strict infinite descending chain C0 > C1 > . . . of centralizers (under inclusion). The class of these groups is denoted by MC . The following fact gives some equivalent reformulations of this notion. Proposition 28.1. Let G be a countable group. Then the following are equivalent: (i) G satisfies the minimal condition on centralizers, (ii) For every increasing sequence G0 ≤ G1 ≤ . . . of subgroups of G, CG (G0 ) ≥ CG (G1 ) ≥ . . . eventually stabilizes, i.e., ∃n∀i ≥ n(CG (Gi ) = CG (Gn )), (iii) Same as (ii) with each Gi finitely generated, (iv) The partial order H1 < H2 of strict inclusion on the set of centralizers is well founded, i.e., for every nonempty set C of centralizers there is a minimal element H0 ∈ C, i.e., H0 is such that for no H1 ∈ C we have H1 < H0 , (v) G satisfies the maximal condition on centralizers, i.e., there is no strict infinite ascending chain C0 < C1 < . . . of centralizers, (vi) For any decreasing sequence G0 ≥ G1 ≥ . . . of subgroups of G, CG (G0 ) ≤ CG (G1 ) ≤ . . . eventually stabilizes, (vii) For every subset A ⊆ G, there is finite B ⊆ A with CG (A) = CG (B).
28. THE MINIMAL CONDITION ON CENTRALIZERS
169
Proof. The equivalence of (i), (ii) is clear since if C0 ≥ C1 ≥ . . . is a decreasing sequence of centralizers, then Ci = CG (Hi ) for some subgroup Hi ≤ G and by letting Gi = hH0 , . . . , Hi i we have G1 ≤ G2 ≤ . . . and CG (Gi ) = Ci . Clearly (i) ⇔ (iv). We next verify that (iv) ⇒ (vii). Given A ⊆ G, let C = {CG (B) : B ⊆ A, B finite} and let CG (B0 ) be a minimal element of C. We claim that CG (B0 ) = CG (A). Otherwise there is g ∈ CG (B0 ) \ CG (A), so that for some a0 ∈ A, ga0 6= a0 g. Then g ∈ CG (B0 ) \ CG (B0 ∪ {a0 }), so CG (B0 ∪ {a0 }) < CG (B0 ), a contradiction. To see that (vii) ⇒ (ii), let S G0 ≤ G1 ≤ . . . and put G∞ = n Gn . Then CG (G∞ ) = CG (A) for some finite A ⊆ G∞ and thus A ⊆ Gn for some n and CG (G∞ ) = CG (Gn ) = CG (Gi ), ∀i ≥ n. Obviously (ii) ⇒ (iii). Assume now (iii). This implies that the strict inclusion order H1 < H2 is well founded on the set of centralizers of finite subsets of G. Since this is all we used in the proof of (iv) ⇒ (vii), it follows that (iii) ⇒ (vii). Thus we have proved the equivalence of (i), (ii), (iii), (iv), (viii). Below write C(A) = CG (A) for simplicity. Then note that A ⊆ B ⇒ C(A) ⊇ C(B) and A ⊆ CC(A), so C(A) ⊇ CCC(A). If g ∈ C(A), then g commutes with CC(A), so g ∈ CCC(A), therefore C(A) ⊆ CCC(A), and thus C(A) = CCC(A), i.e., if B = C(A) is a centralizer, then B = CC(B). So the map B 7→ C(B) is an inclusion reversing involution on the set of centralizers, which proves the equivalence of (i) with (v) and (v) with (vi), and the proof is complete. 2 The class of groups MC is quite extensive. It contains all abelian, linear, finitely generated abelian-by-nilpotent groups and is closed under subgroups, finite products and finite extensions; see, e.g., Bryant [Br]. In particular the free groups Fn (1 ≤ n ≤ ∞) are in MC . Perhaps the easiest examples of groups that fail to satisfy the minimal condition on centralizers are infinite direct sums H0 ⊕ H1 ⊕ . . . of centerless groups Hn . (B) We will next establish a dynamical characterization of these groups that will be the key to our application of this concept in the next section. Let G be a countable group and consider the (diagonal) action of G on GN by conjugation g · {gn } = {ggn g −1 }. Denote by Fc (G) the corresponding equivalence relation {gn }Fc (G){hn } ⇔ ∃g∀n(ggn g −1 = hn ). Then we have the following result. Theorem 28.2 (Kechris). Let G be a countable group. Then the following are equivalent: (i) G satisfies the minimal condition on centralizers. (ii) The equivalence relation Fc (G) is smooth.
170
III. COCYCLES AND COHOMOLOGY
Proof. By 22.3, the equivalence relation Fc (G) is not smooth iff there is {h0n } ∈ GN and {gi } ∈ GN such that gi · {h0n } → {h0n } (in GN with the product topology, G being discrete) but gi · {h0n } = 6 {h0n }, ∀i. So assume Fc (G) is not smooth and let {gi }, {h0n } be as above. Then ∀n∀∞ i(gi h0n gi−1 = h0n ) but ∀i∃n(gi h0n gi−1 6= h0n ). Let An = {h00 , h01 , . . . , h0n }. Then CG (A0 ) ≥S CG (A1 ) ≥ . . . does not stabilize, since ∀n∀∞ i(gi ∈ CG (An )) but ∀i(gi 6∈ CG ( n An )). Conversely, assume G 6∈ MC and let G0 ≤ G1 ≤ . . . be finitely generated subgroups of G with CG (G0 ) > CG (G1 ) > · · · . Say Gi = hh00 , . . . , h0ni i, with n0 < n1 < n2 < . . . Consider then the sequence {h0n } ∈ GN and let gi ∈ CG (Gi ) \ CG (Gi+1 ). Then gi · {h0n }|ni = {h0n }|ni , so gi · {h0n } → {h0n }. Also ∀i∃j ≤ ni+1 (gi h0j gi−1 6= h0j ), thus ∀i(gi · {h0n } 6= {h0n }), so Fc (G) is not smooth. 2 29. Cohomology IV: The E0 -ergodic case (A) First let us recall Proposition 26.8: Proposition 29.1. Let E be an E0 -ergodic equivalence relation and G a countable group. Then B 1 (E, G) ⊆ SMOOTH(E, G). Similarly for actions. Combining this and 27.2 we have the following characterization. Theorem 29.2. Let E be an ergodic equivalence relation. Then the following are equivalent: (i) E is E0 -ergodic. (ii) For all countable groups G, B 1 (E, G) ⊆ SMOOTH(E, G). (iii) For some countable group G 6= {1}, B 1 (E, G) ⊆ SMOOTH(E, G). Similarly for actions. Schmidt [Sc5], 3.3 has proved a similar result for abelian, locally compact Polish groups G. (B) When G is abelian, B 1 (E, G) is a subgroup of the abelian Polish group Z 1 (E, G), so if B 1 (E, G) is closed, the equivalence relation ∼ is closed and so SMOOTH(E, G) = Z 1 (E, G). As we will see later in this section this is not always true if G is not abelian, so we will next investigate under what circumstances one can still derive that SMOOTH(E, G) = Z 1 (E, G). ˜ by Since G can be viewed as a closed subgroup of L(X, µ, G) = G, ˜ on identifying g ∈ G with the constant function x 7→ g, the action of G 1 1 Z (E, G) restricts to the action of G on Z (E, G) given by g · α(x, y) = gα(x, y)g −1 , i.e., conjugation. We will denote by Fc (E, G) the corresponding equivalence relation, which is a subequivalence relation of ∼. Similarly we define Fc (a, G) for a ∈ A(Γ, X, µ). Recall from Section 28, (B) that Fc (G) denotes the equivalence relation of conjugacy on GN and that for equivalence relations E, F on X, Y , resp., we let E ≤B F,
29. COHOMOLOGY IV: THE E0 -ERGODIC CASE
171
if E can be Borel reduced to F , i.e., if there is Borel f : X → Y with xEy ⇔ f (x)F f (y). Proposition 29.3. For any ergodic equivalence relation E on (X, µ) and countable group G, Fc (G) ≤B Fc (E, G)|B 1 (E, G). Similarly for actions. F Proof. Fix a partition X = n Xn , n ≥ 0, where µ(Xn ) = 2−n−1 . Given {gn } ∈ GN , let f{gn } : X → G be defined by ( 1, if x ∈ X0 , f{gn } (x) = gn−1 , if x ∈ Xn , n > 0. Put α{gn } (x, y) = f{gn } (y)f{gn } (x)−1 ∈ B 1 (E, G). We claim that {gn }Fc (G){hn } ⇔ α{gn } Fc (E, G)α{hn } . ⇒: Fix g such that ggn g −1 = hn , ∀n. Then gα{gn } (x, y)g −1 = α{hn } (x, y). ⇐: Let g be such that gα{gn } (x, y)g −1 = α{hn } (x, y), i.e., gf{gn } (y)f{gn } (x)−1 g −1 = f{hn } (y)f{hn } (x)−1 , for xEy. By ergodicity Xn meets (almost) every E-class, so there are xEy such that x ∈ X0 , y ∈ Xn+1 . Then ggn g −1 = hn . 2 I do not know if, conversely, Fc (E, G)|B 1 (E, G) ≤B Fc (G). However we have the following. Proposition 29.4. For any ergodic equivalence relation E on (X, µ) and countable group G, if Fc (G) is smooth, then Fc (E, G) is smooth. Similarly for actions. Proof. Let Γ = {γk }∞ k=0 be a countable group acting in a Borel way on X so that E = EΓX . Assume that Fc (G) is smooth. By 22.3 it is enough to show that if {gn } ∈ GN , α ∈ Z 1 (E, G) and gn · α → α, then for some n, gn · α = α. By going to a subsequence, we can assume that ∀γ∀∞ n(gn α(x, γ · x)gn−1 = α(x, γ · x)), a.e.(x). Define ϕ : X → GN by ϕ(x) = {α(x, γk · x)}∞ k=0 . Then for the conjugacy action of G on GN we have gn · ϕ(x) → ϕ(x), a.e.(x). Since Fc (G) is smooth, by 22.3 again, we have that ∀∞ n(gn · ϕ(x) = ϕ(x)), i.e., ∃N ∀n ≥ N ∀γ(gn α(x, γ · x)gn−1 = α(x, γ · x)), a.e.(x).
172
III. COCYCLES AND COHOMOLOGY
Thus there exists N0 and a set of positive measure A0 ⊆ X (which is therefore an a.e. complete section of E) such that x ∈ A0 ⇒ ∀n ≥ N0 ∀γ(gn α(x, γ · x)gn−1 = α(x, γ · x)). We now claim that ∀n ≥ N0 ∀γ(gn α(x, γ · x)gn−1 = α(x, γ · x)), a.e.(x), therefore ∀∞ n(gn · α = α). Indeed, by discarding a null set, we can assume that for any x there is δ ∈ Γ such that δ −1 · x ∈ A0 . So fix x, γ, n ≥ N0 and let δ −1 · x = y ∈ A0 . Then letting α∗ (γ, x) = α(x, γ · x), we have α∗ (γ, δ · y) = α∗ (γδ, y)α∗ (δ, y)−1 . But gn α∗ (γδ, y)gn−1 = α∗ (γδ, y) and gn α∗ (δ, y)gn−1 = α∗ (δ, y), so gn α∗ (γ, x)gn−1 = (gn α∗ (γδ, y)gn−1 )(gn α∗ (δ, y)−1 gn−1 ) = α∗ (γδ, y)α∗ (δ, y)−1 = α∗ (γ, δ · y) = α∗ (γ, x). 2 Corollary 29.5. For any countable group G, the following are equivalent: (i) G satisfies the minimal condition on centralizers. (ii) Fc (G) is smooth. (iii) For every ergodic equivalence relation E, Fc (E, G) is smooth. (iv) For some ergodic equivalence relation E, Fc (E, G) is smooth. We now have: Theorem 29.6 (Kechris). Let E be an E0 -ergodic equivalence relation and assume G is countable and satisfies the minimal condition on centralizers. Then SMOOTH(E, G) = Z 1 (E, G). Similarly for actions. ˜ α for the action Proof. Fix α ∈ Z 1 (E, G) and consider its stabilizer G 1 ˜ on Z (E, G). By Effros’ Theorem it is enough to show of L(X, µ, G) = G ˜ ˜α → G ˜ α (in G/ ˜ G ˜ α ). By passing to that if gn ∈ G and gn · α → α, then gn G subsequences we can assume that on a measure 1 set in X we have (*)
xEy ⇒ ∀∞ n(gn (y)α(x, y)gn (x)−1 = α(x, y)).
Define the equivalence relation Fc,0 (G) on GN by {fn }Fc,0 (G){hn } ⇔ ∃g∀∞ n(gfn g −1 = hn ). Let also Fk on GN be defined by {fn }Fk {hn } ⇔ ∃g∀n ≥ k(gfn g −1 = hn ). S Then F0 ⊆ F1 ⊆ . . . , Fc,0 (G) = k Fk and clearly Fk is Borel reducible to Fc (G), which is smooth, thus Fk is smooth and Fc,0 (G) is hyperfinite. Now we have by (*) xEy ⇒ {gn (x)}Fc,0 (G){gn (y)},
29. COHOMOLOGY IV: THE E0 -ERGODIC CASE
173
so, since E is E0 -ergodic, there is {h0n } ∈ GN such that {gn (x)}Fc,0 (G){h0n }, a.e.(x), ˜ such that i.e. ∃g∀∞ n(gn (x) = gh0n g −1 ), a.e.(x). So there is g ∈ G (**)
∀∞ n(gn (x) = g(x)h0n g(x)−1 ), a.e.(x).
Put g −1 · α = β, i.e., β(x, y) = g(y)−1 α(x, y)g(x). Then from (∗ ), (∗∗ ) we have, a.e., xEy ⇒ ∀∞ n(h0n β(x, y)(h0n )−1 = β(x, y)), so h0n · β → β. Since, by 29.4, Fc (E, G) is smooth, by 22.3 we have ∀∞ n(h0n · ˜ α ). Let hn = gh0n g −1 ∈ G ˜ α . Then by (∗∗ ) β = β) or ∀∞ n(gh0n g −1 ∈ G ∀∞ n(gn (x) = hn (x)), a.e.(x), so d(gn (x), hn (x)) → 0, where d is the discrete metric on G, and then by ˜ n , hn ) → 0, i.e., d(g ˜ n, G ˜ α ) → 0, so Lebesgue Dominated Convergence d(g ˜ ˜ ˜ ˜ gn Gα → Gα in G/Gα . 2 Combining 27.2 and 29.6 we see that if E is an ergodic equivalence relation and G is a countable group, then we have: (i) If E is not E0 -ergodic, then SMOOTH(E, G) 6= Z 1 (E, G) and ∼ on is not smooth. (ii) If E is E0 -ergodic and G ∈ MC , then SMOOTH(E, G) = Z 1 (E, G) and ∼ is smooth.
Z 1 (E, G)
Thus in particular we have: Corollary 29.7. Let G be a countable group satisfying the minimal condition on centralizers. Then for any ergodic E the following are equivalent: (i) E is E0 -ergodic. (ii) SMOOTH(E, G) = Z 1 (E, G). Similarly for actions. There are examples of E0 -ergodic E in which B 1 (E, G) = Z 1 (E, G), so obviously SMOOTH(E, G) = Z 1 (E, G), but G 6∈ MC . For example, let E be given by a free ergodic action of a property (T) group and let G = G0 ⊕ G1 ⊕ . . . , where Gi are centerless, torsion free and amenable. However in certain situations the smoothness of ∼ on Z 1 (E, G) implies that G ∈ MC . Theorem 29.8 (Kechris). Let a be an E0 -ergodic, weak mixing action of F∞ . Then for any countable G the following are equivalent: (i) G satisfies the minimal condition on centralizers. (ii) SMOOTH(a, G) = Z 1 (a, G). Proof. We have already seen that (i) ⇒ (ii). Assume now that (i) fails. Then Fc (G) is not smooth. But notice that Fc (G) is really the same as the conjugacy relation on Hom(F∞ , G), which by 20.1 can be Borel reduced to ∼
174
III. COCYCLES AND COHOMOLOGY
on Z 1 (a, G). It follows that the latter is not smooth, i.e., SMOOTH(a, G) 6= Z 1 (a, G). 2 I do not know if 29.8 holds as well for Fn , 2 ≤ n < ∞. It is also unclear what is the possible complexity of the equivalence relation ∼ on Z 1 (E, G), when E is E0 -ergodic and G 6∈ MC . (C) We will now see some implications of the previous results to the general problem discussed in Jones-Schmidt [JS] of understanding when for an action a ∈ FRERG(Γ, X, µ) the outer automorphism group Out(Ea ) is Polish or equivalently [Ea ] is closed in N [Ea ]. We have already seen various results in that direction, e.g., this holds for any a ∈ FR(Γ, X, µ) if Γ is not inner amenable (see 9.1). It also holds if a ∈ ERG(Γ, X, µ) and Cµ (Ea ) > 1, even if the action is not free (see 8.1). Jones-Schmidt [JS], 4.5 raise the question of whether there is a dynamical characterization of the closedness of [Ea ] in N [Ea ]. They point out that there appears to be no direct connection between the E0 -ergodicity of a and the closedness of [Ea ] in N [Ea ]. It is easy to see that [Ea ] being closed in N [Ea ] does not imply E0 -ergodicity. (For example, take Γ to be a group that is not inner amenable and does not have property (T) and take a ∈ FRERG(Γ, X, µ) \ E0 RG(Γ, X, µ).) Then they refer to the paper Connes-Jones [CJ1] for an example of a group Γ and a free action a ∈ E0 RG(Γ, X, µ) such that [Ea ] is not closed in N [Ea ]. We present below a variation of this example. Example (Connes-Jones [CJ1]). Let Γ0 be a weakly commutative ICC group and fix {γn } ⊆ Γ0 \ {1} such that ∀γ ∈ Γ0 ∀∞ n(γγn = γn γ). w Consider an ergodic, free action b of Γ0 on (X, µ) which is such that γnb → 1, and [Eb ] is not closed in N [Eb ]. For example, we can take this action to be ∗ the conjugacy shift action of Γ0 on X = 2Γ0 (see the proof of 9.5). Let ∆0 be a non-amenable group and let Γ = Γ0 × ∆0 act on Y = X ∆0 , by letting ∆0 acting by shift and Γ0 acting coordinatewise via b. Call this action a. Since the ∆0 -action is E0 -ergodic, the (Γ0 × ∆0 )-action is E0 -ergodic. One can also easily see that it is free: If (γ0 , δ0 ) · p = p, for a positive measure set of p ∈ X ∆0 , then ∀δ ∈ ∆0 (γ0 · p(δ0−1 δ) = p(δ)). If δ0 6= 1, this clearly holds on a null set of p ∈ X ∆0 . So assume δ0 = 1. Thus ∀δ ∈ ∆0 (γ0 · p(δ) = p(δ)). But if γ0 6= 1, {x : γ0 · x = x} is null, so again this only holds on a null set. Thus γ0 = 1. We will finally show that [Ea ] is not closed in N [Ea ], by showing that γna → 1 in N [Ea ] but clearly γna 6→ 1 uniformly. Since {γn } witnesses the weak commutativity of Γ0 and thus of Γ it is enough, by 6.2, to show that w γna → 1. Again for that it is enough to show that if A1 , . . . , Ak ⊆ X are Borel sets and µk is the product measure on X k , then µk ((γnb (A1 ) × · · · × γnb (Ak ))∆(A1 × · · · × Ak )) → 0. As µ(γnb (Ai )∆Ai ) → 0, ∀i ≤ k, this is clear.
29. COHOMOLOGY IV: THE E0 -ERGODIC CASE
175
This example shows that there is a weakly commutative ICC group Γ and a free, E0 -ergodic action a of Γ on (X, µ) with [Ea ] not closed in N [Ea ]. Note that centerless weakly commutative groups do not have the minimal condition on centralizers. Because ifS{γn } witnesses the weak commutativity of Γ and Γ is in MC , then Γ = n CG ({γk }∞ k=n )) and so, by 28.1, Γ = CG ({γk }∞ ), for some n , i.e., γ ∈ Z(Γ), if k ≥ n0 , a contradiction. 0 k k=n0 We will now show that for ICC groups Γ with the minimal condition on centralizers, E0 -ergodicity of an action a implies the closedness of [Ea ] in N [Ea ]. Thus for “nice” groups Γ, there is indeed a direct implication from E0 -ergodicity to the closedness of [Ea ] in N [Ea ]. For any countable group Γ and action a ∈ A(Γ, X, µ), we denote by α0 the cocycle in Z 1 (a, Γ) given by α0 (γ, x) = γ. We now have the following result. Keep in mind that an action has no nontrivial finite factors (i.e., homomorphisms into actions on non-trivial finite probability spaces) iff every finite index subgroup acts ergodically. Proposition 29.9. Let Γ be a countable group and a ∈ FRERG(Γ, X, µ). Assume that either Γ is ICC or else Γ is centerless and every finite index subgroup of Γ acts ergodically (e.g., a is weak mixing). Then if α0 ∈ SMOOTH(a, Γ), the group [Ea ] is closed in N [Ea ]. Proof. Let E = Ea and let Tn ∈ [E] be such that Tn → 1 in N [E] u in order to show that Tn → 1. Put fn (x) = α0 (x, Tn (x)), so that Tn (x) = fn (x) · x. We have, identifying γ with γ a , u
∀γ ∈ Γ(Tn γTn−1 → γ), i.e., if Aγn = {x : Tn γTn−1 (x) 6= γ · x}, then µ(Aγn ) → 0. Now {x : fn (γ · x)γfn (x)−1 6= γ} ⊆ {x : Tn (γTn−1 (Tn (x))) 6= γ · Tn (x)} = Tn−1 (Aγn ), so ∀γ(µ({x : fn (γ · x)γfn (x)−1 6= γ}) → 0), thus fn · α0 → α0 . ˜ α (= L(X, µ, Γ)α ) = {1}. Indeed if f · α0 = α0 , Next we note that Γ 0 0 −1 i.e., f (γ · x)γf (x) = γ, then f (γ · x) = γf (x)γ −1 , so f : X → Γ is a homomorphism of the Γ-action on X and the conjugacy action on Γ. Thus f∗ µ concentrates on a finite conjugacy class of Γ. Our hypothesis then implies that f∗ µ concentrates on {1}, i.e., f = 1. So, by Effros’ Theorem, since α0 ∈ SMOOTH(a, Γ), we have fn → 1 in ˜ Γ, i.e., µ({x : fn (x) 6= 1}) = µ({x : Tn (x) 6= x}) = δu (Tn , 1) → 1. 2 Theorem 29.10 (Kechris). Let Γ be a countable group satisfying the minimal condition on centralizers and let a ∈ A(Γ, X, µ) be a free, E0 ergodic action of Γ. Assume that either Γ is ICC or else Γ is centerless and every finite index subgroup of Γ acts ergodically (e.g., a is weak mixing). Then [Ea ] is closed in N [Ea ]. Proof. By 29.6 and 29.9.
2
176
III. COCYCLES AND COHOMOLOGY
Recall from Section 9 that it is still an open problem to characterize the ICC groups Γ such that for every a ∈ FRERG(Γ, X, µ), [Ea ] is closed in N [Ea ]. One can ask the “dual” question of characterizing the groups Γ such that for every a ∈ FRERG(Γ, X, µ), [Ea ] is not closed in N [Ea ]. We have already seen in 7.2 that every amenable group Γ satisfies this property, in fact [Ea ] is dense in N [Ea ] in this case. The next result provides the converse, for centerless Γ ∈ MC . Theorem 29.11 (Kechris). Let Γ be a centerless countable group with the minimal condition on centralizers. Then the following are equivalent: (i) Γ is amenable, (ii) ∀a ∈ FRERG(Γ, X, µ)([Ea ] is not closed in N [Ea ]), (iii) If s is the shift action of Γ on 2Γ , then [Es ] is not closed in N [Es ], (iv) ∀a ∈ FRERG(Γ, X, µ)([Ea ] is dense in N [Ea ]), (v) [Es ] is dense in N [Es ]. Proof. (i) ⇒ (ii) - (v) follows from 7.2. If Γ is not amenable, then s is E0 -ergodic and mixing, so, by 29.10, [Es ] is closed in N [Es ], i.e., iii) fails. So (i) - (iii) are equivalent. Moreover, in this case, [Es ] is not dense in N [Es ], else [Es ] = N [Es ], which is absurd as the map x ∈ 2Γ 7→ x ∈ 2Γ (where x(γ) = 1 − x(γ)) is in N [Es ] \ [Es ]. Thus (i), (iv), (v) are equivalent. 2 Remark. It should be pointed out here that if E = R×F is the product of two equivalence relations on (X, µ), (Y, ν), resp., and [F ] is not closed in N [F ], then [E] is not closed in N [E]. This is clear since if Tn ∈ [F ], Tn → 1 in N [F ] but Tn 6→ 1 uniformly and we define Tn0 (x, y) = (x, Tn (y)), then Tn0 ∈ [E], Tn → 1 in N [E] but Tn0 6→ 1 uniformly. It follows that if a group Γ has a free ergodic action a with [Ea ] not closed in N [Ea ], so does any product Γ × ∆, ∆ a countable group. There is a somewhat different version of 29.10 that may be worth stating as it seems to provide some insight concerning the role of weak commutativity in the Connes-Jones counterexample. We use below the following variant of a terminology from Jackson-Kechris-Louveau [JKL]: If E is a countable Borel equivalence relation on a standard Borel space X, we call E measurehyperfinite if for every measure µ on X, E is hyperfinite except perhaps on a µ-null set. Theorem 29.12 (Kechris). Let Γ be a countable group and let a ∈ A(Γ, X, µ) be a free, E0 -ergodic action, in which every finite index subgroup acts ergodically (e.g., a is weak mixing). Assume that Γ is not weakly commutative and Fc,0 (Γ) is measure-hyperfinite. Then α0 ∈ SMOOTH(a, Γ) and thus [Ea ] is closed in N [Ea ]. ˜ α = {1}. We next show that α0 ∈ Proof. As in the proof of 29.9, Γ 0 ˜ SMOOTH(a, Γ). As in the proof of 29.6, we start with a sequence gn ∈ Γ such that, ∀γ∀∞ n(gn (γ · x)γgn (x)−1 = γ), a.e.(x),
29. COHOMOLOGY IV: THE E0 -ERGODIC CASE
177
in order to show that gn → 1. As in that proof, the assumption that a ∈ E0 RG(Γ, X, µ) and Fc,0 (Γ) is measure-hyperfinite implies that there is ˜ such that {h0n } ∈ ΓN and g ∈ Γ ∀∞ n(gn (x) = g(x)h0n g(x)−1 ), a.e.(x), and so ∀γ∀∞ n(gn (γ · x) = g(γ · x)h0n g(γ · x)−1 ), a.e.(x), thus ∀γ∀∞ n(g(γ · x)−1 γg(x) commutes with h0n ), a.e.(x). Put H = {γ ∈ Γ : ∀∞ n(γh0n = h0n γ)} ≤ Γ. So ∀γ(g(γ · x)−1 γg(x) ∈ H), a.e.(x), i.e., ∀γ(g(γ · x)H = γg(x)H), a.e.(x). Put ϕ(x) = g(x)H, so that ϕ : X → Γ/H is a Borel homomorphism of the Γ-action on X into the Γ-action on Γ/H. Then ϕ∗ µ is a Γ-invariant measure on Γ/H, so [Γ : H] < ∞. Since the Γ-action on X has no non-trivial finite factors, we must have Γ = H. Thus since Γ is not weakly commutative, ∀∞ n(h0n = 1), and so ∀∞ n(gn (x) = 1), a.e.(x), and therefore gn → 1. 2 Of course if Γ ∈ MC , then Fc,0 (Γ) is hyperfinite but there are groups Γ for which Fc,0 (Γ) is measure-hyperfinite, which are not in MC , e.g., any amenable group not in MC . Thus the extent of the class of groups Γ for which Fc,0 (Γ) is measure-hyperfinite is not clear. Simon Thomas (private communication) constructed examples of groups Γ for which Fc,0 (Γ) is not measure-hyperfinite. They also satisfy the following additional properties: ICC, inner amenable, not weakly commutative and have cost 1 (and fixed price). (D) We will conclude this section with some further facts related to the last remark in Section 7 and which also give an alternative proof of (a somewhat weaker version of) 29.9. Let a ∈ FR(Γ, X, µ), Γ infinite, and let T ∈ N [Ea ]. Then as in Section 21, (C) we associate to each T ∈ N [Ea ] the cocycle αT ∈ Z 1 (a, Γ) defined by T (γ · x) = αT (γ, x) · T (x). Proposition 29.13. The map T 7→ αT is continuous from (N [Ea ], τN [Ea ] ) into Z 1 (a, Γ). Proof. Put E = Ea . Assume that Tn → T in N [E] in order to show w that αTn → αT in Z 1 (a, Γ). Note that Tn → T in N [E] implies that Tn → T w w and so Tn−1 → T −1 and thus Tn−1 T → 1.
178
III. COCYCLES AND COHOMOLOGY
Fix > 0, γ ∈ Γ. We will find N such that ∀n ≥ N , µ({x : αTn (γ, x) 6= αT (γ, x)}) < . n Put Aδ = {x : αT (γ, x) = δ}, P Aδ = {x : αTn (γ, x) = δ}. Then there is a finite F ⊆ Γ such that δ6∈F µ(Aδ ) < /6. Now {x : αTn (γ, x) 6= S S αT (γ, x)} = δ (Aδ ∩ δ0 6=δ Anδ0 ). Choose M such that ∀n ≥ M , ∀δ ∈ F (µ(Aδ ∆Tn−1 T (Aδ )) < ). 6|F | w
This is possible, since Tn−1 T → 1. Then {x : αTn (γ, x) 6= αT (γ, x)} =
[
(Aδ ∩
δ∈F
Anδ0 )∪
δ 0 6=δ
δ6∈F
[
[
(Aδ ∩
[
Anδ0 ),
δ 0 6=δ 0
and {x : αTn (γ, x) 6= αT (γ, T −1 Tn (x)} =
[
(Tn−1 T (Aδ ) ∩
Anδ0 )∪
δ 0 6=δ
δ6∈F
[
[
(Tn−1 T (Aδ ) ∩
δ∈F
[
Anδ0 ),
δ 0 6=δ 0
so µ({x : αTn (γ, x) 6= αT (γ, x)}∆{x : αTn (γ, x) 6= αT (γ, T −1 Tn (x))}) < /6 + /6 + We also have that for n ≥ N ,
Tn γTn−1
→
X
µ(Aδ ∆Tn−1 T (Aδ )) < /2·
δ∈F T γT −1
uniformly, so we can find N > M such
µ({x : Tn γTn−1 (x) 6= T γT −1 (x)}) < /2. By the freeness of the action, {x : αTn (γ, Tn−1 (x)) 6= αT (γ, T −1 (x))} ⊆ {x : Tn γTn−1 (x) 6= T γT −1 (x)}, so µ({x : αTn (γ, Tn−1 (x)) 6= αT (γ, T −1 (x))}) < /2. Therefore µ({x : αTn (γ, Tn−1 (x)) 6= αT (γ, T −1 (x))}) = µ({x : αTn (γ, x) 6= αT (γ, T −1 Tn (x))}) < /2, so we have µ({x : αTn (γ, x) 6= αT (γ, x)}) < . 2
30. COHOMOLOGY V: ACTIONS OF PROPERTY (T) GROUPS
179
Recall that for an action a ∈ A(Γ, X, µ), Ca = {T ∈ Aut(X, µ) : T aT −1 = a} is the stabilizer of the action (also denoted by CΓ if the action is understood). Then we have: Proposition 29.14 (Furman [Fu]). For any a ∈ FR(Γ, X, µ), T ∈ N [Ea ], T ∈ Ca [Ea ] ⇔ αT ∼ α0 . Proof. ⇒: If T = SU , where S ∈ Ca , U ∈ [Ea ], and f : X → Γ is such that f (x) · x = U (x), then U (γ · x) = f (γ · x)γf (x)−1 · U (x), so T (γ · x) = S(U (γ · x)) = S(f (γ · x)γf (x)−1 · U (x)) = f (γ · x)γf (x)−1 · T (x), so αT (γ, x) = f (γ · x)γf (x)−1 , i.e., αT ∼ α0 . ⇐: Suppose that αT (γ, x) = f (γ · x)γf (x)−1 , for some f ∈ L(X, µ, Γ). Then T (γ · x) = f (γ · x)γf (x)−1 · T (x). Put S(x) = f (x)−1 · T (x), so that S(γ · x) = γ · S(x). We claim that S ∈ Aut(X, µ). First notice that S maps any orbit Γ · x in a 1-1 way onto the orbit Γ · S(x) = Γ · T (x) and thus, in particular, S is a Borel bijection of X. Next we will verify that S preserves µ, and for that it is enough to show that for each δ ∈ Γ, if Aδ = f −1 ({δ}), then S|Aδ is measure preserving. But for x ∈ Aδ , S(x) = δ −1 · T (x), so this is clear. Thus S ∈ Ca . Put U (x) = f (S −1 (x)) · x. Then U S = T , so U = T S −1 ∈ Aut(X, µ). Moreover U ∈ [Ea ]. So T ∈ [Ea ]Ca = Ca [Ea ]. 2 From 29.13 and 29.14, it follows that if a ∈ FR(Γ, X, µ) and the cocycle α0 ∈ SMOOTH(a, Γ), then Ca [Ea ] is closed in N [Ea ], so by the last remark in Section 7, if moreover Ca ∩ [Ea ] = {1}, [Ea ] is closed in N [Ea ]. Recall from 14.6 that if a ∈ FRERG(Γ, X, µ), then Ca ∩ [Ea ] = {1} if either Γ is ICC or else Γ is centerless and every infinite subgroup of Γ acts ergodically. Compare this with 29.9. 30. Cohomology V: Actions of property (T) groups (A) For actions of property (T) groups the structure of the cohomology classes is much simpler in view of the following result. Theorem 30.1 (Popa [Po3]). Let Γ be a countable group that has property (T), let G be a countable group and let a ∈ ERG(Γ, X, µ). Then every cohomology class in Z 1 (a, G) is clopen (so H 1 (a, G) is countable). Proof (Furman [Fu2]). We will find δ > 0 such that if α, β ∈ Z 1 (a, G) and d˜1 (α, β) < δ, then α ∼ β. Since Γ has property (T) there is a finite symmetric F ⊆ Γ and > 0 such that for any unitary representation π : Γ → U (H), which admits a unit vector v0 with ∀γ ∈ F (|hπ(γ)(v0 ), v0 i| ≥ 1 − ), there is a Γ-invariant unit 1 vector v ∈ H with kv − v0 k < 10 . 1 Fix now α, β ∈ Z (a, G) and consider the following action of Γ on X ×G: γ · (x, g) = (γ · x, α(γ, x)gβ(γ, x)−1 ).
180
III. COCYCLES AND COHOMOLOGY
Clearly the σ-finite measure µ × ηG , where ηG is the counting measure on G, is Γ-invariant, and this gives rise to the unitary representation π : Γ → U (L2 (X × G, µ × ηG )), given as usual by π(γ)(v)(x, g) = v(γ −1 · (x, g)). Consider the vector v0 = χX×{1} . Clearly kv0 k = 1. Also for γ ∈ Γ, Z −1 hπ(γ )(v0 ), v0 i = χX×{1} (γ · (x, g))χX×{1} (x, g)d(µ × ηG ) = µ({x : α(γ, x) = β(γ, x)}). Choose now δ > 0 such that d˜1 (α, β) < δ ⇒ ∀γ ∈ F (µ({x : α(γ, x) 6= β(γ, x)}) < ), and thus for γ ∈ F , d˜1 (α, β) < δ ⇒ 1 ≥ |hπ(γ)(v0 ), v0 i| ≥ 1 − . It follows that if d˜1 (α, β) < δ, there is a Γ-invariant unit vector v with 1 kv − v0 k < 10 . We will use this to find f : X → G such that f · β = α, thus α ∼ β. Now v ∈ L2 (X × G, µ × ηG ) and v(x, g) = v(γ · x, α(γ, x)gβ(γ, x)−1 ), ∀γ. Put fx (g) = |v(x, g)|2 . Then fγ·x (α(γ, x)gβ(γ, x)−1 ) = fx (g) and g 7→ α(γ, x)gβ(γ, x)−1 is a permutation of G, so the maps X V (x) = fx (g), g∈G
m(x) = maxg∈G fx (g) and k(x) = card{g ∈ G : fx (g) = m(x)} are Γ-invariant. By the ergodicity of the Γ-action on X, these functions are constant, a.e. Note that, by Fubini, V (x) = V = 1. Also m(x) = m is positive and ∞ > k(x) = k > 0. If we can show that k(x) = 1, then letting f (x) = (the unique g such that fx (g) = m), we have f (γ · x) = α(γ, x)f (x)β(γ, x)−1 , i.e., f · β = α. But if k ≥ 2, then |v(x, 1)|2 ≤ 12 , a.e., so Z 2 kv − v0 k2 = |v(x, g) − v0 (x, g)|2 d(µ × ηG ) Z 1 2 1 2 > , ≥ |v(x, 1) − 1| dµ ≥ ( 1 − √ 100 2 a contradiction.
2
The preceding and earlier results lead to the following characterization of property (T) groups.
30. COHOMOLOGY V: ACTIONS OF PROPERTY (T) GROUPS
181
Theorem 30.2. Let Γ be a countable group. Then the following are equivalent: (i) Γ has property (T), (ii) For every a ∈ ERG(Γ, X, µ) and any countable G, the cohomology classes are clopen in Z 1 (a, G), (iii) For every a ∈ ERG(Γ, X, µ), there is some Polish G 6= {1} admitting an invariant metric, such that SMOOTH(a, G) = Z 1 (a, G), (iv) For every a ∈ ERG(Γ, X, µ) and any countable G, H 1 (a, G) is countable, (v) For every a ∈ ERG(Γ, X, µ), there is some Polish G 6= {1} admitting an invariant metric with H 1 (a, G) countable. Proof. (i) ⇒ (ii) is 30.1 and (ii) ⇒ (iii) is trivial. If (i) fails, then Γ has an ergodic but not E0 -ergodic action a and so, by 27.2, (iii) fails. Finally (ii) ⇒ (iv) ⇒ (v) ⇒ (iii) are clear. 2 A special case of this characterization, for abelian G, was proved in Schmidt [Sc3], 3.4. Also note that, as we discussed earlier in Section 27, (B), if Γ has property (T) and G is a torsion-free group with HAP, then B 1 (a, G) = Z 1 (a, G) for any a ∈ ERG(Γ, X, µ). Remark. There is an analog of 30.1 valid for any countable group Γ, brought to my attention by Ioana; see Ioana [I2]. It is reminiscent of the last remark of Section 14, (A). Let a ∈ A(Γ, X, µ), G a countable group with discrete metric d, and put on Z 1 (a, G) the complete metric d˜1∞ (α, β) = sup
Z d(α(γ, x), β(γ, x))dµ(x)
γ∈Γ
= sup µ({x : α(γ, x) 6= β(γ, x)}). γ∈Γ
Then for a ∈ ERG(Γ, X, µ), every cohomology class in Z 1 (a, G) is clopen in the d˜1∞ -topology. Indeed, following the proof of 30.1, we note that for γ ∈ Γ, 1−hπ(γ)(v0 ), v0 i ≤ d˜1∞ (α, β), so kπ(γ)(v0 )−v0 k2 ≤ 2d˜1∞ (α, β). If v1 is the unique element of least norm in the closed convex hull of the Γ-invariant set {π(γ)(v0 ) : γ ∈ Γ}, then v1 is Γ-invariant and kv1 − v0 k ≤ 2d˜1∞ (α, β). Thus by choosing d˜1∞ (α, β) sufficiently small, we can find a Γ-invariant unit vector with kv − v0 k < 1/10 and repeating the rest of the proof of 30.1, we conclude that α ∼ β. We can also see that for Γ with property (T), for a ∈ ERG(Γ, X, µ) and G countable, the d˜1∞ -topology on Z 1 (a, G) is the same as the usual topology on Z 1 (a, G). This follows from the proof of 30.1.
182
III. COCYCLES AND COHOMOLOGY
1 First notice that if kv − v0 k2 < ρ < 10 , then Z ρ2 > kv − v0 k22 = |v(x, g) − χX×{1} (x, g)|2 d(µ × ηG ) Z XZ 2 |v(x, g)| dµ(x) + |v(x, 1) − 1|2 dµ(x) = g6=1
≥
X
1 4 µ({x
: |v(x, g)|2 ≥ 14 }) + 14 µ({x : |v(x, 1) − 1|2 ≥ 14 })
g6=1
so 4ρ2 > µ({x : ∃g 6= 1(|v(x, g)|2 ≥ 12 ) or |v(x, 1) − 1| ≥ 12 }), therefore µ({x : ∀g 6= 1(|v(x, g)| < 21 ) and |v(x, 1) − 1| < 12 }) > 1 − 4ρ2 , thus µ({x : f (x) = 1}) > 1 − 4ρ2 . We also have f · β = α, that is, f (γ · x)β(γ, x)f (x)−1 = α(γ, x), ∀γ, so µ({x : α(γ, x) 6= β(γ, x)}) ≤ 8ρ2 , ∀γ, i.e., d˜1∞ (α, β) ≤ 8ρ2 . 1 Now given 10 > ρ > 0 we can choose δ > 0 such that if d˜1 (α, β) < δ, then there is v with kv − v0 k2 < ρ, so d˜1∞ (α, β) ≤ 8ρ2 . This shows that the above two topologies coincide. It also follows from 30.2 and 27.2 that the following are equivalent: (i) Γ has property (T), (ii) For every a ∈ ERG(Γ, X, µ) and countable G, the d˜1∞ -topology on 1 Z (a, G) coincides with the usual topology. (iii) For every a ∈ ERG(Γ, X, µ) and countable G, d˜1∞ is separable. (B) One can use 30.1 to derive superrigidity results. First we need the following characterization. Theorem 30.3 (Popa [Po3]). Suppose Γ is a countable group, G a Polish group admitting an invariant metric and a ∈ WMIX(Γ, X, µ), α ∈ Z 1 (a, G). Consider the product action a × a ∈ A(Γ, X 2 , µ2 ) and let α1 (γ, (x1 , x2 )) = α(γ, x1 ), α2 (γ, (x1 , x2 )) = α(γ, x2 ), 1 so that αi ∈ Z (a × a, G), i = 1, 2. Then the following are equivalent: (i) α is cohomologous to a group homomorphism ϕ : Γ → G, (ii) α1 ∼ α2 . Proof (Furman [Fu2]). (i) ⇒ (ii): If α ∼ ϕ, let f : X → G be such that ϕ(γ) = f (γ ·x)α(γ, x)f (x)−1 . Then if f1 (x1 , x2 ) = f (x1 ) we have f1 ·α1 = ϕ, so α1 ∼ ϕ. Similarly α2 ∼ ϕ, thus α1 ∼ α2 . (ii) ⇒ (i): Fix f : X 2 → G with α(γ, x2 ) = f (γ · x1 , γ · x2 )α(γ, x1 )f (x1 , x2 )−1 , µ2 -a.e.(x1 , x2 ). It follows that
30. COHOMOLOGY V: ACTIONS OF PROPERTY (T) GROUPS
183
α(γ, x2 ) = f (γ · x1 , γ · x2 )α(γ, x1 )f (x1 , x2 )−1 , α(γ, x2 ) = f (γ · x3 , γ · x2 )α(γ, x3 )f (x3 , x2 )−1 , µ3 -a.e.(x1 , x2 , x3 ). So α(γ, x1 ) = f (γ · x1 , γ · x2 )−1 f (γ · x3 , γ · x2 )α(γ, x3 )f (x3 , x2 )−1 f (x1 , x2 ). Put g(x1 , x2 , x3 ) = f (x1 , x2 )−1 f (x3 , x2 ). Then for the action a3 = a × a × a ∈ A(Γ, X 3 , µ3 ), we have g(γ · (x1 , x2 , x3 )) = α(γ, x1 )g(x1 , x2 , x3 )α(γ, x3 )−1 , µ3 -a.e.(x1 , x2 , x3 ). Claim. g depends only on x1 , x3 , i.e., there is g : X 2 → G such that g(x1 , x2 , x3 ) = g(x1 , x3 ) µ3 -a.e.(x1 , x2 , x3 ). Granting this, we also have g(γ · x1 , γ · x2 , γ · x3 ) = g(γ · x1 , γ · x3 ), µ3 -a.e.
(x1 , x2 , x3 ), so g(γ · x1 , γ · x3 ) = α(γ, x1 )g(x1 , x3 )α(γ, x3 )−1 ,
µ3 -a.e. (x1 , x2 , x3 ), i.e., g(γ · x1 , γ · x3 ) = α(γ, x1 )g(x1 , x3 )α(γ, x3 )−1 , µ2 -a.e., (x1 , x3 ). Fix x02 such that g(x1 , x02 , x3 ) = f (x1 , x02 )−1 f (x3 , x02 ) = g(x1 , x3 ), µ2 -a.e. (x1 , x3 ), and thus g(γ · x1 , γ · x3 ) = f (γ · x1 , x02 )−1 f (γ · x3 , x02 ), µ2 -a.e. (x1 , x3 ). Put h(x) = f (x, x02 ). Then α(γ, x1 ) = h(γ · x1 )−1 h(γ · x3 )α(γ, x3 )h(x3 )−1 h(x1 ), µ2 -a.e.(x1 , x3 ), so h(γ · x1 )α(γ, x1 )h(x1 )−1 = h(γ · x3 )α(γ, x3 )h(x3 )−1 , µ2 -a.e.(x1 , x3 ). It follows that ϕ(γ) = h(γ · x)α(γ, x)h(x)−1 is constant µ-a.e.(x). Clearly then ϕ is a homomorphism and α ∼ ϕ. Proof of the claim. We have for an invariant metric d on G and every γ, d(g(γ · (x1 , x2 , x3 )), g(γ · (x1 , x4 , x3 ))) = d(g(x1 , x2 , x3 ), g(x1 , x4 , x3 )) µ4 -a.e.(x1 , x2 , x3 , x4 ), so (x1 , x2 , x3 , x4 ) 7→ d(g(x1 , x2 , x3 ), g(x1 , x4 , x3 ))
184
III. COCYCLES AND COHOMOLOGY
is invariant for the action a4 = a × a × a × a ∈ A(Γ, X 4 , µ4 ). Since a is weak mixing, a4 is ergodic, so the above function is constant, a.e., say with value d0 . We will show that d0 = 0, i.e., g(x1 , x2 , x3 ) = g(x1 , x4 , x3 ) µ4 -a.e.(x1 , x2 , x3 , x4 ). Then choose x04 so that g(x1 , x2 , x3 ) = g(x1 , x04 , x3 ), µ3 a.e.(x1 , x2 , x3 ), and put g(x1 , x3 ) = g(x1 , x04 , x3 ). Assume d0S> 0, towards a contradiction, and fix {gn0 } dense in G. Then we have that n B(gn0 , d0 /2) = G. Fix x01 , x03 such that d(g(x01 , x2 , x03 ), g(x01 , x4 , x03 )) = d0 , µ2 -a.e.(x2 ,S x4 ). Let q(x) = g(x01 , x, x03 ). Then if Bn = q −1 (B(gn0 , d0 /2)), we have that n Bn = X, so for some n, µ(Bn ) > 0. Thus µ2 (Bn2 ) > 0, so fix x02 , x04 ∈ Bn such that d(g(x01 , x02 , x03 ), g(x01 , x04 , x03 )) = d0 , therefore d(q(x02 ), q(x04 )) = d0 , contradicting the fact that d(q(x02 ), gn0 ), d(q(x04 ), gn0 )) <
d0 2 .
2 a2
Following Popa [Po3] we call a ∈ A(Γ, X, µ) malleable if for = a× a ∈ A(Γ, X 2 , µ2 ), there is a continuous path in Ca2 = {T ∈ Aut(X 2 , µ2 ) : T a2 T −1 = a2 } from 1 to the flip (x, y) 7→ (y, x). Note now the following fact. Proposition 30.4. Let Γ be a countable group and let a ∈ WMIX(Γ, X, µ) be malleable. Let G be a Polish group with an invariant metric and consider H 1 (a2 , G) = Z 1 (a2 , G)/∼ with the quotient topology. If H 1 (a2 , G) is totally path disconnected (i.e., its path components are trivial), then a is G-superrigid. Proof. Fix α ∈ Z 1 (a, G) and consider α1 , α2 as in 30.3. Let t 7→ Tt ∈ Ca2 connect 1 with the flip. Now Ca2 acts continuously on Z 1 (a2 , G) (see Section 24, (B)), so Tt · α1 is a continuous path in Z 1 (a2 , G) from α1 to α2 . Projecting to H 1 (a2 , G), we conclude that α1 ∼ α2 , so α is cohomologous to a homomorphism by 30.3. 2 We thus have: Theorem 30.5 (Popa [Po3]). Let Γ have property (T) and let the action a ∈ WMIX(Γ, X, µ) be malleable. Then for any countable G, a is G-superrigid. Proof. By 30.1, clearly H 1 (a2 , G) is discrete.
2
A canonical example of a malleable action of a group Γ is its shift s on [0, 1]Γ , where [0, 1] has Lebesgue measure λ. If we identify T ∈ Aut([0, 1]2 , λ2 ) with T ∗ ∈ Aut(([0, 1]2 )Γ , (λ2 )Γ ), given by T ∗ (p) = (γ 7→ T (p(γ))), then the group Aut([0, 1]2 , λ2 ) becomes a closed subgroup of Cs2 .
30. COHOMOLOGY V: ACTIONS OF PROPERTY (T) GROUPS
185
The flip on [0, 1]Γ × [0, 1]Γ = ([0, 1]2 )Γ is equal to T ∗ , where T is the flip on [0, 1]2 . Since Aut([0, 1]2 , λ2 ) is path connected, there is a continuous path in Cs2 from 1 to the flip, so s is malleable. The previous superrigidity result can be extended as follows. Theorem 30.6 (Popa [Po3]). Let Γ be a countable group containing a normal subgroup ∆ Γ with relative property (T). Let a ∈ Aut(Γ, X, µ) be malleable such that a|∆ ∈ WMIX(∆, X, µ). Then a is G-superrigid for every countable G. Proof. First note that for any b ∈ ERG(Γ, Y, ν) the proof of 30.1 shows that if α, β ∈ Z 1 (b, G) and d˜1 (α, β) < δ, for an appropriate δ, then α|∆ ∼ β|∆. Then the proof of 30.3 shows that if α ∈ Z 1 (a, G) and α1 |∆ ∼ α2 |∆, then α|∆ ∼ ϕ for some ϕ ∈ Hom(∆, G). As before the malleability of a implies that there is a path in Z 1 (a2 , G) from α1 to α2 and thus by applying the preceding observation to b = a2 ∈ ERG(Γ, X 2 , µ2 ), we see (by cutting this path into small pieces) that α1 |∆ ∼ α2 |∆, so α|∆ ∼ ϕ, for some ϕ ∈ Hom(∆, G). We will now use that ∆ Γ to find ψ ∈ Hom(Γ, G) such that α ∼ ψ. We have α(δ, x) = f (δ · x)ϕ(δ)f (x)−1 , ∀δ ∈ ∆, for some f ∈ L(X, µ, G). Put β = f −1 · α, so that β|∆ = ϕ. Put Γ0 = {γ ∈ Γ : β(γ, x) = β(γ) is constant a.e.}. Clearly ∆ ⊆ Γ0 . We will show that Γ0 = Γ, which will complete the proof. Fix γ ∈ Γ, δ ∈ ∆. Then γδγ −1 = δ 0 ∈ ∆, so β(γδ, x) = β(γ, δ · x)β(δ, x) = β(γ, δ · x)β(δ) = β(δ 0 γ, x) = β(δ 0 , γ · x)β(γ, x) = β(δ 0 )β(γ, x). Thus β(γ, δ · x)β(δ) = β(γδγ −1 )β(γ, x) and so β(γ, δ · x) = β(γδγ −1 )β(γ, x)β(δ)−1 . Therefore for each γ ∈ Γ, Aγ = {(x, y) ∈ X 2 : β(γ, x) = β(γ, y)} is ∆invariant for the action a2 , so since a|∆ is weak mixing, and thus a2 |∆ is ergodic, µ2 (Aγ ) = 0 or 1. If for all γ it has measure 1, we are done by Fubini. Otherwise for some γ, β(γ, x) 6= β(γ, y), µ2 -a.e.(x, y), contradicting the fact that for some g0 , B = {x : β(γ, x) = g0 } has positive measure, thus µ2 (B) > 0. 2 Here is a sample application of 30.5 to orbit equivalence. Theorem 30.7 (Popa [Po3]). Let Γ be a simple countable group with property (T), and let s be the shift action of Γ on X = [0, 1]Γ . Let ∆ be a countable group and a ∈ FR(∆, Y, ν). If sOEa, then there is an isomorphism ϕ : Γ → ∆ and an isomorphism q : (X, µ) → (Y, ν) such that q(γ · x) = ϕ(γ) · q(x), i.e., the groups Γ, ∆ are isomorphic and identifying Γ, ∆ via an isomorphism the actions s, a are isomorphic.
186
III. COCYCLES AND COHOMOLOGY
Proof. Fix an orbit equivalence p : X → Y, xEs y ⇔ p(x)Ea p(y), and let α : Γ × X → ∆ be the corresponding cocycle: α(γ, x) · p(x) = p(γ · x). By 30.5 (and the paragraph following it), there is a homomorphism ϕ : Γ → ∆ such that α ∼ ϕ, so fix Borel f : X → ∆ such that f (γ · x)α(γ, x)f (x)−1 = ϕ(γ). Then if q(x) = f (x) · p(x), we have that xEs y ⇔ q(x)Ea q(y) and q(γ · x) = ϕ(γ) · q(x). Since Γ is simple, ϕ is an injection and thus q is an injection. Let Aδ = f −1 ({δ}). Then {Aδ }δ∈∆ is a partition of X and q(x) = δ · p(x) for x ∈ Aδ , so q|Aδ is measure preserving and thus so is q. Thus q : (X, µ) → (Y, ν) is an isomorphism and the proof is complete. 2 (C) We conclude with some remarks and questions concerning path connectedness in H 1 (a, G). In general it seems interesting to understand under what conditions H 1 (a, G) is totally path disconnected. First, let us note that if Γ is finitely generated, a ∈ E0 RG(Γ, X, µ) and G is a countable group in MC , then if a is G-superrigid, H 1 (a, G) is totally path disconnected. To see this notice that there are only countably many homomorphisms from Γ to G, thus H 1 (a, G) is countable, while by 29.6 every point in H 1 (a, G) is closed. Thus any path in H 1 (a, G) must be constant by Sierpinski’s Theorem (see Section 27, (G)). I do not know if finite generation of Γ is necessary here. Second, consider a strongly treeable countable group Γ, i.e., a group Γ such that for every a ∈ FR(Γ, X, µ), Ea is treeable (µ-a.e.) (see KechrisMiller [KM], Section 30). There are many such groups, e.g., Γ = Z2 ∗ Z3 , which admit no non-trivial homomorphism into F2 . Since for any a ∈ FR(Γ, X, µ), Ea is treeable, it can be Borel reduced to the free part of the equivalence relation induced by the shift of F2 on 2F2 (see JacksonKechris-Louveau [JKL], 3.5), so the corresponding reduction cocycle is not cohomologous to a homomorphism and thus a is not F2 -superrigid. It follows from 30.4 that if Γ is strongly treeable, with Hom(Γ, F2 ) trivial, and s is the shift action of Γ on [0, 1]Γ (which is malleable and isomorphic to s × s), then H 1 (s, F2 ) is not totally path disconnected. In fact the preceding arguments also show the following. Proposition 30.8. Let Γ be a non-amenable, finitely generated group, G a countable group in MC and let s be the shift action of Γ on [0, 1]Γ . Then the following are equivalent: i) s is G-superrigid, ii) H 1 (s, G) is totally path disconnected.
APPENDIX A
Realifications and complexifications We will consider below real and complex Hilbert spaces and some connections between them. I would like to thank Scot Adams for a useful conversation on these matters. If H is a real Hilbert space, we define its complexification HC as follows: The elements of HC are the formal sums x+i·y (so, as a set, HC = H 2 ) with vector operations defined in the obvious way: (x1 + i · y1 ) + (x2 + i · y2 ) = (x1 + x2 ) + i · (y1 + y2 ), (a + bi)(x + i · y) = (ax − by) + i · (bx + ay), for a, b ∈ R. The inner product is again defined by hx1 + i · y1 , x2 + i · y2 iHC = hx1 , x2 iH + hy1 , y2 iH + ihy1 , x2 iH − ihx1 , y2 iH . We will write HC = H + i · H, and identify x ∈ H with x + i · 0. The norm of x + i · y in HC is kx + i · yk2HC = hx + i · y, x + i · yiHC = hx, xiH + hy, yiH + ihy, xiH − ihx, yiH = kxk2H + kyk2H . It follows that HC is a complex Hilbert space. If {eα } is an orthonormal basis in H, then {eα } is also an orthonormal basis for HC . If H is a complex Hilbert space, we define its realification HR by restriction of scalars. The elements of HR are exactly those of H and the vector operations are defined by restriction to R. The inner product is given by hx, yiHR = Re hx, yiH . Thus kxk2HR = Re hx, xiH = hx, xiH = kxk2H , so that the norms are the same and HR is a real Hilbert space. If {eα } is an orthonormal basis for H, then {eα , ieα } is an orthonormal basis for HR . If H is a real Hilbert space, the identity map is a Hilbert space embedding of H into (HC )R (i.e., a real Hilbert space isomorphism of H with a closed subspace on (HC )R ). If H is a complex Hilbert space, then we can define a canonical Hilbert space embedding of H into (HR )C as follows: 1 ϕ(x) = √ (x + i · (−ix)) 2 If H is a real Hilbert space we denote by O(H) the orthogonal group of H, i.e., the group of Hilbert space isomorphisms of H. We equip O(H) with the strong operator topology, i.e., the one induced by the maps T 7→ T (x) from O(H) to H for all x ∈ H. This is the same as the weak operator topology, i.e., the one induced by the maps T 7→ hT (x), yi from O(H) to R 187
188
A. REALIFICATIONS AND COMPLEXIFICATIONS
for all x, y ∈ H. If H is separable, with this topology, O(H) is a Polish group. If H is a complex Hilbert space, we similarly define the unitary group of H, U (H). If H is a real Hilbert space and T ∈ O(H), let TC = T + i · T be the element of U (HC ) given by (T + i · T )(x + i · y) = T (x) + i · T (y). Clearly TC is the unique T1 ∈ U (HC ) with T1 |H = T and {TC : T ∈ O(H)} = {T1 ∈ U (H) : T1 (H) = H}, so via T 7→ TC (which is easily a topological group isomorphism) we can identify O(H) with the closed subgroup of U (HC ) consisting of elements that leave H invariant, so O(H) ⊆ U (HC ). If H is a complex Hilbert space and T ∈ U (H), clearly T ∈ O(HR ) and the identity map is a topological group isomorphism of U (H) onto a closed subgroup of O(HR ), thus we have again U (H) ⊆ O(HR ). If H is a real Hilbert space and we view H, via the identity map, as a closed subspace of (HC )R , then, via T 7→ T + i · T, O(H) is identified with a closed subgroup of O((HC )R ), so O(H) ⊆ O((HC )R ). If H is a complex Hilbert space, T ∈ U (H) and we identify H with a ˜ of (HR )C via x 7→ √1 (x + i · (−ix)), then every T ∈ U (H) closed subspace H 2 ˜ defined by T˜( √1 (x + i · (−ix)) = √1 (T (x) + i · gives rise to T˜ ∈ U (H) 2
2
(−iT (x))). But also T gives rise to T + i · T ∈ U ((HR )C ). We note that ˜ to (HR )C . Thus if we identify T ∈ U (H) with T + i · T extends T˜ from H ˜ invariant and (T + i · T )|H ˜ = T˜. T + i · T ∈ U ((HR )C ), then T + i · T leaves H ˜ via the identification T 7→ T˜ Thus we can view U (H) as either U (H) or as a closed subgroup of U ((HR )C ) via the identification T 7→ T + i · T . ˜ is The above shows that these identifications cohere, in the sense that H ˜ gives T˜. invariant under each T + i · T and its restriction to H
APPENDIX B
Tensor products of Hilbert spaces Below we consider either real or complex vector and Hilbert spaces. If V1 , . . . , Vn are vector spaces, their (algebraic) tensor product is the (unique up to isomorphism) vector space V1 ⊗ · · · ⊗ Vn equipped with a multilinear map V1 × · · · × Vn → V1 ⊗ · · · ⊗ Vn , (v1 , . . . , vn ) 7→ v1 ⊗ · · · ⊗ vn such that for any vector space V and multilinear map π : V1 × · · · × Vn → V there is unique linear map ρ : V1 ⊗ · · · ⊗ Vn → V with π(v1 , . . . , vn ) = ρ(v1 ⊗ · · · ⊗ vn ). Every element of V1 ⊗ · · · ⊗ Vn is a linear combination of vectors of the form v1 ⊗ · · · ⊗ vn . If B1 , . . . , Bn are bases for V1 , . . . , Vn , resp., then B1 ⊗ · · · ⊗ Bn = {v1 ⊗ · · · ⊗ vn : vi ∈ Bi } is a basis for V1 ⊗ · · · ⊗ Vn . Now assume that H1 , . . . , Hn are Hilbert spaces. Then there is a unique inner product on the algebraic tensor product H1 ⊗ · · · ⊗ Hn such that hv1 ⊗ · · · ⊗ vn , w1 ⊗ · · · ⊗ wn iH1 ⊗···⊗Hn =
n Y hvi , wi iHi . i=1
Then we define the Hilbert space tensor product, also denoted by H1 ⊗ · · · ⊗ Hn as the completion of this inner product space. (We have used H1 ⊗ · · · ⊗ Hn to denote both the algebraic and the Hilbert space tensor product. Since it is the latter we are interested in, we will use this notation only for the latter in the sequel.) Note that the linear combinations of the vectors v1 ⊗ · · · ⊗ vn , vi ∈ Hi are dense in H1 ⊗ · · · ⊗ Hn . Also kv1 ⊗ · · · ⊗ vn kH1 ⊗···⊗Hn = kv1 kHi · · · kvn kHn . Finally, if B1 , . . . , Bn are orthonormal bases for H1 , . . . , Hn , resp., then B1 ⊗ · · · ⊗ Bn is an orthonormal basis for H 1 ⊗ · · · ⊗ Hn . If H1 = · · · = Hn = H, we let H ⊗n = H1 ⊗ · · · ⊗ Hn . By convention ⊗0 H = the scalar field (viewed as a 1-dimensional Hilbert space) and H ⊗1 = H. We next define the symmetric tensor power of H. Note that for each σ ∈ Sn (= the permutation group of {1, . . . , n}) there is a unique operator on H ⊗n also denoted by σ such that σ(v1 ⊗ · · · ⊗ vn ) = vσ(1) ⊗ · · · ⊗ vσ(n) . Let H n be the (closed) subspace of all vectors in H ⊗n invariant under all σ ∈ Sn . Also H n can be viewed as the range (or equivalently the P set of fixed 1 P √1 σ. Let v · · · v = points) of the operator τ = n! 1 n σ∈Sn σ∈Sn vσ(1) ⊗ n! √ · · ·⊗vσ(n) = n!τ (v1 ⊗· · ·⊗vn ). Then the linear combinations of the vectors n vP 1 · · · v Qnn , vi ∈ H, are dense in H . Moreover hv1 · · · vn , w1 · · · wn i = σ∈Sn i=1 hvi , wσ(i) i, so that if B is an orthonormal basis for H, then 189
190
B. TENSOR PRODUCTS OF HILBERT SPACES
B · · · B is an orthonormal basis for H n . We finally have H 0 = scalars, H 1 = H. The direct sum H ∞ = H 0 ⊕ H 1 ⊕ · · · ⊕ H n ⊕ . . . is called the symmetric Fock space or boson Fock space of H. Any bounded linear map L : H1 → H2 between Hilbert spaces extends to a unique bounded linear map L⊗n : H1⊗n → H2⊗n , so that L⊗n (v1 ⊗ · · · ⊗ vn ) = L(v1 ) ⊗ · · · ⊗ L(vn ) and similarly L n : H1 n → H2 n so that L n (v1 · · · vn ) = L(v1 ) · · · L(vn ). L n , We have kL⊗n k = kL n k = kLkn . Finally, if kLk ≤ 1, let L ∞ = ∞ n=0 L which is a bounded linear map from H1 ∞ to H2 ∞ . In particular, If T is an orthogonal (resp., unitary) operator on a real (resp. complex) Hilbert space H, then T induces a unique orthogonal (resp., unitary) operator T ⊗n on H ⊗n so that T ⊗n (v1 ⊗ · · · ⊗ vn ) = T (v1 ) ⊗ · · · ⊗ T (vn ), and similarly T n on H n so that T n (v1 · · · vn ) = T (v1 ) · · · T (vn ). L n on H ∞ . The map T 7→ T ∞ is Finally let T ∞ be the operator ∞ n=0 T an isomorphism of O(H) (resp., U (H)) with a closed subgroup of O(H ∞ ) (resp., U (H ∞ )). Clearly T ∞ |H = T . Comments. We have followed here the exposition of tensor products in Janson [Ja], Appendix E.
APPENDIX C
Gaussian probability spaces Suppose F is a finite (non-∅) set and ϕ : F 2 → R a real positivedefiniteP function on F , i.e., a symmetric function (ϕ(i, j) = ϕ(j, i)), satisfying i,j∈F ai aj ϕ(i, j) ≥ 0 for all reals ai , i ∈ F . Then a standard result in probability theory asserts that there is a unique probability Borel measure µϕ on RF such that the sequence of projection functions pi (x) = xi , where x = (xi )i∈F , from RF to R is a Rd -valued (where d = card(F )) Gaussian random variable with mean 0 and covariance matrix ϕ. This means that the following two conditions are satisfied: P F → R is a centered (i) Every linear combination f = i∈F αi pi : R Gaussian random variable, i.e., has centered Gaussian distribution N (0, σ 2 ), for some σ. Explicitly this means that f∗ µϕ is the centered Gaussian measure x2
1 with density √2πσ e− 2σ2 (−∞ < x < ∞), σ ≥ 0. When σ = 0, we interpret N (0, 0) as the Dirac (point) measure at 0. R (ii) Cor(pi pj ) = E(pi pj ) = xi xj dµϕ = ϕ(i, j), for (i, j) ∈ F 2 .
Note that the projection of µϕ to the ith coordinate is the centered Gaussian measure N (0, ϕ(i, i)). P characteristic function of this measure, R The which is defined by µ ˜ϕ (u) = ei( k∈F uk xk ) dµϕ (x), µ ˜ϕ : RF → C, is given by the following formula 1
µ ˜ϕ (u) = e− 2
P
i,j∈F
ui uj ϕ(i,j)
.
Recall that µ ˜ϕ uniquely determines µϕ . We call µϕ the Gaussian centered measure associated to ϕ. For example, if ϕ(i, j) = δij (= the Kronecker delta), then µϕ is the product measure on RF , where in each coordinate we have the standard Gaussian measure with distribution N (0, 1) (i.e., density x2 √1 e− 2 ). 2π
Now consider a countable set I. Using the preceding and the Kolmogorov Consistency Theorem, we see P that for each real positive-definite ϕ : I 2 → R (i.e., ϕ(i, j) = ϕ(j, i), i,j∈F ai aj ϕ(i, j) ≥ 0, for all finite F ⊆ I and reals ai , i ∈ F ), there is a unique probability Borel measure µϕ on RI such that for each finite (non-∅) F ⊆ I, if πF : RI → RF is the canonical projection πF (x) = x|F (x ∈ RI ), then (πF )∗ µϕ = µϕ|F . ThusPµϕ is uniquely determined by saying that each linear combination f = i∈F αi pi : RI → R, where pi (x) = xi , is a centered Gaussian random 191
192
C. GAUSSIAN PROBABILITY SPACES
R variable and Cor(pi pj ) = E(xi xj ) = xi xj dµϕ = ϕ(i, j). We call this the Gaussian centered measure associated to ϕ. Again if ϕ(i, j) = δi,j , µϕ is the product measure µI , where µ = N (0, 1). A special case arises as follows: Let Γ be a countable group and ϕ : Γ → R a real positive-definite function on Γ, i.e., one for which the function ϕ0 : Γ2 → R given by ϕ0 (γ, δ) = ϕ(γ −1 δ) −1 and for any finite F ⊆ Γ is positive definite. This means that Pϕ(γ) = ϕ(γ )−1 and any reals aγ , γ ∈ F , we have γ,δ∈F aγ aδ ϕ(γ δ) ≥ 0. Consider then the measure µϕ0 on RΓ and the shift action of Γ on RΓ : γ · f (δ) = f (γ −1 δ). Using the characterization of µϕ0 in terms of the properties of the projection function pγ (x) = xγ , it is easy to check that this action preserves µϕ0 . For simplicity we will write µϕ instead of µϕ0 when ϕ : Γ → R is positive-definite. We will call the shift action of Γ on (RΓ , µϕ ) the Gaussian shift (associated with ϕ). Comments. Some references here are Jacod-Protter [JP], §16, Glasner [Gl2], pp. 89–91, Khoshnevisan [Kh], §5.
APPENDIX D
Wiener chaos decomposition (A) Let I be a countable set, ϕ a real positive-definite function ϕ : → R and consider the Gaussian space (RI , µϕ ). There is a canonical decomposition of L2 (RI , µϕ , R) (the real Hilbert space of real-valued square integrable functions on (RI , µϕ )) called the Wiener chaos decomposition. Let H = the closed linear span of the projection functions pi (x) = xi , i ∈ I. Then it is not hard to see that every f ∈ H has centered Gaussian distribution (i.e., f∗ µϕ has distribution N (0, σ 2 ), for some σ 2 ). (Such closed linear subspaces of L2 are called (real) Gaussian Hilbert spaces.) Next for each n ≥ 0, let P¯n (H) be the closure in L2 (RI , µϕ , R) of the linear subspace Pn (H) consisting of all p(f1 , . . . , fm ), p a real polynomial of degree ≤ n (in m variables for some m), and f1 , . . . , fm ∈ H. Then P¯0 (H) = R (the constant functions) ⊆ P¯1 (H) ⊆ P¯2 (H) ⊆ P¯3 (H) ⊆ . . . and we let (following standard notation) H :0: = R, H :n: = P¯n (H) P¯n−1 (H) = P¯n (H) ∩ P¯n−1 (H)⊥ , n ≥ 1.
I2
Then (noticing that H⊥R as the projection functions have mean 0) we have that H :1: = H, {H :n: }n≥0 are pairwise orthogonal and (using that the projectionSfunctions separate points, so they generate the Borel algebra of RI ) that n Pn (H) = P (H) is dense in L2 (RI , µϕ , R). Therefore M L2 (RI , µϕ , R) = H :n: . n≥0
This is called the Wiener chaos decomposition of L2 (RI , µϕ , R). H = H :1: is called the first chaos. (B) This decomposition has an alternate description using symmetric tensor products. There is a canonical Hilbert space isomorphism of H n with H :n: . This isomorphism sends f1 · · · fn (where fi ∈ H) to the so-called Wick product of f1 , . . . , fn denoted by : f1 . . . fn : which is by definition the projection of the product f1 . . . fn into the space H :n: . Thus under this identification of H n with H :n: , we can also write M M H ∞ = H n = H :n: = L2 (RI , µϕ , R). n≥0
n≥0
(C) Suppose now ϕ, ψ are real positive-definite functions and we denote by Hϕ , Hψ the corresponding first Wiener chaos for ϕ, ψ, resp. Let 193
194
D. WIENER CHAOS DECOMPOSITION
T : Hϕ → Hψ be a Hilbert space isomorphism. Then, since T extends canonically to the symmetric Fock space of Hϕ (see Appendix B), there is a canonical isomorphism T ∞ : Hϕ ∞ = L2 (RI , µϕ , R) → Hψ ∞ = L2 (RI , µψ , R) extending T . We will actually see that there is a (unique) isomorphism S : (RI , µϕ ) → (RI , µψ ) which gives T in the sense that the corresponding isomorphism OS : L2 (RI , µϕ , R) → L2 (RI , µψ , R) induced by S, i.e., OS (f ) = f ◦S −1 , is equal to T ∞ , i.e., T ∞ (f ) = f ◦S −1 , f ∈ L2 (RI , µϕ , R). To prove this, consider the projection functions pi (x) = xi on RI viewed as elements of Hϕ and let T (pi ) = qi ∈ Hψ . Then let θ : RI → RI be given by θ(x)i = qi (x). We claim that θ is 1-1 on a µψ -measure 1 set and θ∗ µψ = µϕ . Thus θ : (RI , µψ ) → (RI , µϕ ) is an isomorphism. Let S = θ−1 . Then OS (pi )(x) = pi (θ(x)) = θ(x)i = qi (x), i.e., OS (pi ) = qi , so OS |Hϕ = T and, since OS preserves multiplication, it follows that OS = T ∞ . To see that θ is 1-1 on a µψ -measure 1 set, note that the closed linear span of {qi } contains each pi , thus the σ-algebra generated by {qi } is the Borel σ-algebra modulo µψ -null sets, so there is a Borel µψ -conull set A such that {qi |A} generates the Borel σ-algebra of A and thus {qi |A} separates points and therefore θ is 1-1 on A. We will finally verify that θ∗ µψ = µϕ . For this it is enough to check that (RI , θ∗ µψ ) is the Gaussian centered measure space associated to ϕ. Translated back to (RI , µψ ) this means that every linear combination f = Pn i=1 αi qi has centered Gaussian distribution, which is clear as f ∈ Hψ , and that Eµψ (qi qj ) = ϕ(i, j), which is clear as Eµψ (qi qj ) = hqi , qj iHψ = hT (pi ), T (pj )iHψ = hpi , pj iHϕ = Eµϕ (pi pj ) = ϕ(i, j). In particular, if H is the first chaos of L2 (RI , µϕ , R) and T ∈ O(H) = the orthogonal group of H, then T ∞ ∈ O(H ∞ ) = O(L2 (RI , µϕ , R)) extends T canonically to L2 (RI , µϕ , R) and there is a unique S ∈ Aut(RI , µϕ ) so that if OS ∈ O(L2 (RI , µϕ , R)) is the associated orthogonal operator given by OS (f ) = f ◦ S −1 , then OS |H = T and OS = T ∞ . Thus identifying T and S we can view O(H) as the closed subgroup of Aut(RI , µϕ ) (with the weak topology) consisting of all S ∈ Aut(RI , µϕ ) for which OS (H) = H. Note that the complex Hilbert space L2 (RI , µϕ , C) is in the obvious sense the complexification of L2 (RI , µϕ , R) L2 (RI , µϕ , C) = L2 (RI , µϕ , R) + i · L2 (RI , µϕ , R). Also we have the decomposition (under again some obvious identifications) L L :n: :n: HC ∞ = H ∞ + i · H ∞ = n≥0 HC:n: = n≥0 (HR + i · HR ) 2 I = L (R , µϕ , C). Moreover if we define Pn,C (H) by using complex polynomials of degree ≤ n, then we have HC:n: = Pn,C (H) Pn−1,C (H), where the closure is now in L2 (RI , µϕ , C). (Note also that HC:0: = C.)
D. WIENER CHAOS DECOMPOSITION
195
(D) Now consider the special case where Γ is a countable group and ϕ : Γ → R is a real positive-definite function, and the corresponding shift action of Γ on (RΓ , µϕ ). Let p1 (x) = x1 be the projection to the identity element of Γ (p1 : RΓ → R). Considering the orthogonal Koopman representation of Γ on L2 (RΓ , µϕ , R) associated to the shift (i.e., γ · f (x) = f (γ −1 · x)), we have γ · p1 (x) = p1 (γ −1 · x) = (γ −1 · x)1 = xγ = pγ (x), i.e., γ · p1 = pγ . Thus H = hpγ iγ∈Γ = hγ · p1 iγ∈Γ , where hXi denotes the closed linear span of X. Moreover, hγ · p1 , p1 i = hpγ , p1 i = E(pγ p1 ) = ϕ(γ −1 ) = ϕ(γ) (as every real positive-definite function is symmetric). Recall that given a real positive-definite function ϕ on Γ there is a unique (up to the obvious notion of isomorphism) triple (Hϕ , πϕ , eϕ ) consisting of a real Hilbert space Hϕ , an orthogonal representation πϕ of Γ on Hϕ , and a cyclic vector eϕ ∈ Hϕ (i.e., hπϕ (γ)(eϕ )iγ∈Γ = Hϕ ) such that ϕ(γ) = hπϕ (γ)(eϕ ), eϕ i. This is the well-known GNS construction. It follows that the first Wiener chaos H with the restriction of the orthogonal representation induced by the shift action and p1 is the GNS representation associated to ϕ. (E) Suppose now that ϕ1 is the characteristic function of {1} ⊆ Γ. Clearly this is positive-definite. The corresponding µϕ1 is simply the product µΓ , where µ = normalized Gaussian measure on R = N (0, 1), so we are looking at the shift action of Γ on (RΓ , µΓ ). Then {pγ }γ∈Γ clearly is an orthonormal basis for H on which Γ acts transitively: γ · pδ = pγδ . So the orthogonal representation of Γ on H is simply (isomorphic to) the (left-) regular representation of Γ on `2 (Γ, R). For each finite multiset F ⊆ Γ of size n ≥ 2, i.e., an object of the form P F = {n1 γ1 , . . . , nk γk }, where ni ≥ 1, ki=1 nk = n, γ1 , . . . , γk ∈ Γ distinct, let p F = pγ1 · · · pγ1 · · · pγk · · · pγk , where γi is repeated ni times. Then for each n ≥ 2, {p F : size of F = n} is an orthonormal basis for H n . Denote by (Γ)n the set of all multisets of size n. Then Γ acts on (n) (n) (Γ)n by γ · {n1 γ1 , . . . , nk γk } = {n1 γγ1 , . . . , nk γγk }. Let O1 , O2 , . . . be (n) (n) (n) (n) the orbits of this action. Choose Fi ∈ Oi and let Γi = {γ : γ · Fi = (n) (n) Fi }, which is a finite subgroup of Γ. Then clearly hp F : F ∈ Oi i = (n) Hi is a closed subspace of H n invariant under the orthogonal Koopman (n) representation and the representation of Γ on Hi is isomorphic to the (n) (n) quasi-regular representation of Γ on `2 (Γ/Γi , R) (given by γ · f (δΓi ) = (n) f (γ −1 δΓi )). Thus the orthogonal representation of Γ on L L2 (RΓ , µΓ , R) associated to the shift action is isomorphic to 1Γ,R ⊕ λΓ,R ⊕ n,i λΓ/Γ(n) ,R i where 1Γ,R is the trivial 1-dimensional orthogonal representation, λΓ,R is the regular orthogonal representation on `2 (Γ, R) and λΓ/∆,R is the quasi-regular orthogonal representation on `2 (Γ/∆, R). Complexifying we have that if κ is
196
D. WIENER CHAOS DECOMPOSITION
the unitary Koopman representation (on L2 (RΓ , µΓ , C)) of the shift action of Γ, then M κ∼ λΓ/Γ(n) , = 1Γ ⊕ λΓ ⊕ n,i
i
where we are now referring to unitary representations (on the spaces C, (n) (n) `2 (Γ, C), `2 (Γ/Γi , C), resp.). If Γ is torsion-free, so that Γi = {1}, then κ∼ = 1Γ ⊕ ∞ · λΓ . (where ∞ · ρ means the sum of ℵ0 many copies of ρ). In general, if κ0 = κ|L20 (RΓ , µΓ , C)), so that κ = 1Γ ⊕ κ0 , it can be shown that κ0 ≤ ∞ · λΓ (see Bekka-de la Harpe-Valette [BdlHV], E.4.5). Comments. For the Wiener chaos decomposition and Gaussian Hilbert spaces, see Janson [Ja].
APPENDIX E
Extending representations to actions Let H be a separable real Hilbert space of dimension 1 ≤ N ≤ ∞. Consider the product space (RN , µN ), where µ is the standard Gaussian measure with distribution N (0, 1). Then if {pi } are the projection functions, pi (x) = xi , the first Wiener chaos H :1: = hpi i has dimension N , as {pi } is an orthonormal basis for it. Fix an isomorphism θ : H ∼ = H :1: . Consider now T ∈ O(H) and θT θ−1 ∈ O(H :1: ). Then there is unique S ∈ Aut(RN , µN ) such that OS |H :1: = θT θ−1 . We can identify T with S and simply view O(H) as a closed subgroup of Aut(RN , µN ). (This identification of course depends on θ.) It follows that an orthogonal representation π of a countable group Γ on H gives rise to an action aπ ∈ A(Γ, RN , µN ). Identifying H and H :1: via θ, it is clear that π is the restriction of the orthogonal Koopman representation associated with aπ to the space H :1: . Fix now a separable complex Hilbert space H. Let HR be its realification and let N be the dimension of HR (as a real Hilbert space). Consider again the product space (RN , µN ). Fix an isomorphism θ : HR ∼ = H :1: . Then θ :1: ∼ extends to an isomorphism, θC : (HR )C = HC of the complexifications of HR , H :1: . Consider now T ∈ U (H). Then T ∈ O(HR ) and θT θ−1 ∈ O(H :1: ). Thus there is unique S ∈ Aut(RN , µN ) such that OS |H :1: = θT θ−1 . We can identify again T with S and simply view U (H) as a closed subgroup of Aut(RN , µN ). It follows that a unitary representation π ∈ Rep(Γ, H) of a countable group Γ on H gives rise to an action aπ ∈ A(Γ, RN , µN ). It clearly has the property that if π 0 is the orthogonal Koopman representation on L2 (RN , µN , R) induced by aπ , π 0 leaves H :1: invariant and π 0 |H :1: is (via the identification via θ) the orthogonal representation induced by π on HR . Clearly if π ∼ = ρ (as unitary representations), then aπ ∼ = aρ (as actions). Next consider the unitary Koopman representation κaπ induced by aπ on L2 (RN , µN , C) (thus κaπ = π 0 + i · π 0 ). Let also κa0π be its restriction to L20 (RN , µN , C). We claim that π ≤ κa0π , i.e., π is isomorphic to a subrepresentation of κa0π . To see this, recall from Appendix A that there is a copy ˜ of H in (HR )C , so that the copy T˜ of any T ∈ U (H) is the restriction of H ˜ It follows that θC (H) ˜ ⊆ H :1: ⊆ L2 (RN , µN , C) is invariant T + i · T to H. 0 C aπ aπ ∼ ˜ under κ0 and that κ0 |θC (H) = π, so π ≤ κa0π . The discussion in this appendix provides a proof of the following result stated in Zimmer [Zi1], 5.2.13.
197
198
E. EXTENDING REPRESENTATIONS TO ACTIONS
Theorem E.1. Let Γ be a countable group and π a unitary representation of Γ on a (complex) Hilbert space H. Then one can explicitly construct a measure preserving action aπ of Γ on (X, µ) such that π ∼ = aρ and = ρ ⇒ aπ ∼ aπ 2 if κ0 is the unitary Koopman representation of Γ on L0 (X, µ) associated to aπ , then π ≤ κa0π . If π has no non-0 finite-dimensional subrepresentations, i.e., it is weak mixing, then aπ is weak mixing. This is proved by an argument as in the proof of 11.2. See also Glasner [Gl2], 3.59.
APPENDIX F
Unitary representations of abelian groups ˆ its dual group (A) Let ∆ be a countable abelian group and denote by ∆ ˆ is a compact Polish group. Denote (where ∆ is viewed as discrete). So ∆ ˆ ˆ = δ(δ) ˆ ˆ where ˆ by hδ, δi the duality between ∆ and ∆, so that δ(δ) = hδ, δi, ˆ and δˆ ∈ ∆ ˆ as a character of ∆. Let we view δ ∈ ∆ as a character of ∆ ϕ : ∆ → C be a positive-definite function, i.e., for any finite F ⊆ ∆ and P −1 aγ ∈ C, γ ∈ F , we have a γ,δ∈F γ aδ ϕ(γ δ) ≥ 0. Then, by Bochner’s Theorem (see, e.g., Rudin [Ru2]), there is a unique finite (positive) Borel ˆ such that for any δ ∈ ∆, if we put measure µ on ∆ Z µ ˆ(δ) = δdµ, then µ ˆ = ϕ. Conversely, for each such µ, µ ˆ is positive definite. Let now π ∈ Rep(∆, H) be a unitary representation of ∆ on a separable Hilbert space H. For each h ∈ H, let ϕπ,h (δ) = ϕh (δ) = hπ(δ)(h), hi. Then ϕh is positive-definite, so let µπ,h = µh be the corresponding measure ˆ on ∆: µ ˆπ,h (δ) = µ ˆh (δ) = hπ(δ)(h), hi. A vector h0 ∈ H is called a cyclic vector for π if H is the closed linear span of {π(δ)(h0 ) : δ ∈ ∆}. If π admits a cyclic vector h0 , then (H, π, h0 ) is the GNS representation associated with ϕh0 . For each finite Borel measure µ ˆ consider also the representation πµ of ∆ on L2 (∆, ˆ µ) given by: on ∆ πµ (δ)(f ) = δf. ˆ µh ), πµ , 1) is also the GNS representation corresponding to Then (L2 (∆, 0 h0 ϕh0 , so there is an isomorphism of π with πµh0 sending h0 to 1. Now recall that a projection valued measure (into H) on a standard Borel space X is a map E : Bor(X) → Pr(H) from the class of Borel subsets of X into the set of projections on HPsuch that E(∅) = 0, E(X) = I, E(A ∩ B) = S E(A)E(B) and E( ∞ A )= ∞ i i=1 i=1 E(Ai ), if {Ai } are pairwise disjoint. 199
200
F. UNITARY REPRESENTATIONS OF ABELIAN GROUPS
ˆ → We now claim that there is a projection valued measure E : Bor(∆) ˆ Pr(H) such that for any h ∈ H and Borel set A ⊆ ∆, µh (A) = hE(A)(h), hi = kE(A)(h)k2 . Such an E is called a spectral measure for the representation π. Since we can decompose H into a direct sum of cyclic subrepresentations (i.e., subrepresentations admitting a cyclic vector), we can assume that π admits a cyclic ˆ µ), h0 = 1, π = πµ , vector h0 and therefore we can assume that H = L2 (∆, ˆ → Pr(L2 (∆, ˆ µ)) such that where µ = µh0 . So we need to define E : Bor(∆) 2 ˆ µf (A) = hE(A)(f ), f i, for each f ∈ L (∆, µ). Indeed, let ˆ µ) E(A) = χA L2 (∆, ˆ µ) supported by A. Then be the projection to the subspace Rof all f ∈ L2 (∆, R we need toR check that µf (A) = χRA dµf = hχA f, f i = χA |f |2 dµ. Now for δ ∈ ∆, δdµf = hπµ (δ)(f ), f i = δ|f |2 dµ, so, using the fact that linear ˆ are uniformly dense in C(∆) ˆ (by Stonecombinations of characters of ∆ Weierstrass) and using the Lebesgue Dominated Convergence Theorem, we R R ˆ → C we have gdµf = g|f |2 dµ, so see that for any bounded Borel g : ∆ we are done putting g = χA . ˆ (B) The maximal spectral type for π is a finite Borel measure σ on ∆ ˆ such that for any Borel set A ⊆ ∆: σ(A) = 0 ⇔ E(A) = 0 ⇔ ∀h ∈ H(µh (A) = 0). This is uniquely determined up to measure equivalence. Fix an orthonormal basis {en }∞ n=1 for H and put σπ =
∞ X 1 µe . 2n n
n=1
Then clearly σπ is the maximal spectral type of π and the map π 7→ σπ is ˆ the space of probcontinuous from Rep(∆, H) (see Section 13) into P (∆), ∗ ˆ ability Borel measures on ∆ with the weak -topology. Note that if h0 is a cyclic vector for π, then µh0 is the maximal spectral type for π. To ˆ µ) see this, it is enough to consider the representation π = πµ on L2 (∆, ˆ µ), and the cyclic Rvector 1, whose measure is µ. If f ∈ L2 (∆, R associated 2 2 then µ ˆf (δ) = (δf )f dµ = δ|f | dµ = νˆ(δ), where dν = |f | dµ. Thus µf = ν µ. It follows that for any π there is h ∈ H with µh the maximal spectral type of π. (C) Now let ∆ be infinite and consider the Gaussian shift s on R∆ corresponding to ϕ1 = the characteristic function of {1} ⊆ ∆. The corresponding Gaussian measure is the product measure µ∆ , where µ is the normalized Gaussian measure on R. Thus s is free and ergodic. We will next calculate that the maximal spectral type for κs0 is η∆ ˆ = Haar measure
F. UNITARY REPRESENTATIONS OF ABELIAN GROUPS
201
ˆ (see Section 10, (C) for the definition of the Koopman representation on ∆ s κ0 corresponding to s). Using Appendix D, (E) and noticing the obvious fact that η∆ ˆ is the maximal spectral type of λ∆ = the left-regular representation of ∆, we see that it is enough to show that if F ≤ ∆ is a finite subgroup and λ∆/F is the quasi-regular representation given by the action of ∆ on `2 (∆/F ), then the maximal spectral type for λ∆/F is η∆ ˆ . Inˆ deed, let Λ ≤ ∆ be the intersection of the kernels of all δ ∈ F . As F is ˆ is finite, any such δ ∈ F satisfies δ n = 1, for some n, and therefore δ(∆) finite, so ker(δ) has finite index and thus is clopen. So Λ is clopen and its η |Λ Haar measure is ηΛ = η ˆ∆ˆ(Λ) . The measure σ associated to the cyclic vector ∆
χ{F } ∈ `2 (∆/F ) of the representation λ∆/F is the maximal spectral type of λ∆/F and satisfies σ ˆ (δ) = 1, if δ ∈ F, σ ˆ (δ) = 0, if δ 6∈ F . Clearly ηˆΛ (δ) = 1, if δ ∈ F . If δ 6∈ F , then by Rudin [Ru2], proof of 2.1.3, there is δˆ ∈ Λ such ˆ 6= 1. So δ|Λ is a non-trivial character of Λ, thus ηˆΛ (δ) = 0. So that δ(δ) σ ˆ (δ) = ηˆΛ (δ), for all δ, and thus σ = ηΛ η∆ ˆ.
APPENDIX G
Induced representations and actions We will review here some facts about induced representations and actions. Our general reference is Bekka-de la Harpe-Valette [BdlHV], Part II. (A) Let Γ ≤ ∆ be countable groups. Fix a transversal T for the left cosets of Γ in ∆ containing 1 ∈ Γ. Then ∆ acts on T by δ · t = the unique element of T in the coset δtΓ, and we also define the following cocycle for this action, ρ : ∆ × T → Γ, ρ(δ, t) = γ, where δt = (δ · t)γ. Let now π ∈ L Rep(Γ, H) be a unitary representation of Γ on H. Consider the direct sum t∈T H of card(T ) = |T | copies of H. Define then the induced representation M Ind∆ H) Γ (π) ∈ Rep(∆, t∈T
by M M δ·( ht ) = ρ(δ −1 , t)−1 · hδ−1 ·t . t∈T
t∈T
It is easy to check, using the cocycle property for ρ, that this is indeed a unitary representation. Let now a ∈ A(Γ, X, µ), and assume that [∆ : Γ] < ∞, i.e., T is finite. Denote the normalized counting measure on T by νT . Consider the space X × T with measure µ × νT and define the induced action Ind∆ Γ (a) ∈ A(∆, X × T, µ × νT ) by δ · (x, t) = (ρ(δ, t) · x, δ · t). It is easy to check that this is indeed measure preserving. If a is free (resp., ergodic), then Ind∆ Γ (a) is free (resp., ergodic). We next verify that for [∆ : Γ] < ∞ and a ∈ A(Γ, X, µ), ∆ a κIndΓ (a) ∼ = Ind∆ Γ (κ ),
i.e., induction commutes with the Koopman representation. To see this let ∆ f ∈ L2 (X × T, µ × νT ) = the Hilbert space of κIndΓ (a) , and define U (f ) ∈ 203
204
L
t∈T
G. INDUCED REPRESENTATIONS AND ACTIONS a L2 (X, µ) = the Hilbert space of Int∆ Γ (κ ), by 1 M ft , U (f ) = p |T | t∈T
where ft (x) = f (x, t). Then it can be easily checked L that U is an isomorphism of the Hilbert spaces L2 (X ×T, µ×νT ) and t∈T L2 (X, µ). We finally ∆ a verify that it preserves the ∆-actions (corresponding to κIndΓ (a) , Ind∆ Γ (κ ), resp.). Indeed, we have U (δ · f ) = U ((x, t) 7→ f (δ −1 · (x, t))) = U ((x,L t) 7→ f (ρ(δ −1 , t) · x, δ −1 · t))) 1 −1 −1 · t)). = √ t (x 7→ f (ρ(δ , t) · x, δ |T |
On the other hand, L δ · U (f ) = δ · √1 t ft |T | L −1 −1 · f = √1 δ −1 ·t t ρ(δ , t) |T | L 1 −1 = √ t (x 7→ fδ −1 ·t (ρ(δ , t) · x)) |T | L −1 −1 · t)), = √1 t (x 7→ f (ρ(δ , t) · x, δ |T |
so U (δ · f ) = δ · U (f ). It follows that if a ∈ A(Γ, X, µ), b ∈ A(Γ, Y, ν) are unitarily equivalent, ∆ i.e., κa ∼ = κb , then so are Ind∆ Γ (a), IndΓ (b). ˆ We use below (B) Now assume that ∆ is abelian with dual group ∆. the notation of Appendix F. If θ is an automorphism of ∆, then we define ˆ δ)i ˆ = hθ(δ), δi. ˆ ˆ by hδ, θ( the automorphism of θˆ of ∆ For each representation π ∈ Rep(∆, H) and automorphism θ of ∆, we have the corresponding representation θ(π) = π ◦ θ ∈ Rep(∆, H). For each δ ∈ ∆, h ∈ H, µ ˆθ(π),h (δ) = hθ(π)(δ)(h), hi = hπ(θ(δ))(h), hi = ˆ ∗ µπ,h and therefore if σ is the maxiϕπ,h (θ(δ)) = µ ˆπ,h (θ(δ)), so µθ(π),h = (θ) ˆ mal spectral type for π, (θ)∗ σ is the maximal spectral type for θ(π). Let now ∆ Γ with [Γ : ∆] < ∞, and L consider π ∈ Rep(∆, H) and the Γ inducedL representation Ind∆ (π) ∈ Rep(Γ, t∈T H). Let π = IndΓ∆ (π)|∆ ∈ Rep(∆, t∈T H). Then for the representation π we have L L δ · ( t∈T ht ) = Lt∈T ρ(δ −1 , t)−1 · hδ−1 ·t = Lt∈T (t−1 δt) · ht , = t∈T θt (δ) · ht , where θt (δ) = t−1 δt. Thus π ¯=
M t∈T
θt (π).
G. INDUCED REPRESENTATIONS AND ACTIONS
205
It follows that if σ is the maximal spectral type for π, then X (θˆt )∗ σ t∈T
is the maximal spectral type for π ¯ = IndΓ∆ (π)|∆. (I would like to thank Dinakar Ramakrishnan for a useful discussion here.)
APPENDIX H
The space of unitary representations (A) We use here Bekka-de la Harpe-Valette [BdlHV] as a general reference for unitary representations. Let H be a separable complex Hilbert space. We endow U (H), the unitary group of H, with the strong topology, i.e., the one generated by the family of maps T ∈ U (H) 7→ T (x) ∈ H, x ∈ H. This is the same as the weak topology, generated by the maps T ∈ U (H) 7→ hT (x), yi ∈ C, x, y ∈ H. Then U (H) is a Polish group. If now Γ is a countable group, we let Rep(Γ, H) be the set of all homomorphisms for Γ into U (H), i.e., the set of unitary representations of Γ on H. We also write H = Hπ if π ∈ Rep(Γ, H). If we give U (H)Γ the product topology, so it becomes a Polish space, then Rep(Γ, H) is a closed subspace of U (H)Γ , thus it is Polish too. If {em } is an orthonormal basis for H, then an open nbhd basis for π ∈ Rep(Γ, H) consists of the sets of the form {σ ∈ Rep(Γ, H) : ∀i, j ≤ k∀γ ∈ F (|hσ(γ)(ei ), ej i − hπ(γ)(ei ), ej i| < )}, where k = 1, 2, . . . , F ⊆ Γ is finite and > 0. The group U (H) acts continuously on Rep(Γ, H) by conjugation, i.e., T · π = T πT −1 , where T πT −1 (γ) = T π(γ)T −1 . Let us note here that if FN is the free group with N generators, N ∈ {1, 2, . . . , N} and α1 , α2 , . . . are free generators, the map π ∈ Rep(FN , H) 7→ (π(α1 ), π(α2 ), . . . ) ∈ U (H)N is a homeomorphism of Rep(FN , H) with the space U (H)N preserving conjugation. Thus we can identify Rep(FN , H) with U (H)N . In particular, Rep(Z, H) can be identified with U (H). The representations π ∈ Rep(Γ, Hπ ), ρ ∈ Rep(Γ, Hρ ) are isomorphic, in symbols π∼ = ρ, if there is an isomorphism T of Hπ , Hρ with T π(γ)T −1 = ρ(γ), ∀γ ∈ Γ. If we consider the conjugacy action of U (H) on Rep(Γ, H), then clearly for π, ρ ∈ Rep(Γ, H), π ∼ = ρ iff π, ρ are conjugate. We have the following easy fact. Proposition H.1. Let Γ be a countable group. If H is infinite-dimensional, there is a dense conjugacy class in Rep(Γ, H). 207
208
H. THE SPACE OF UNITARY REPRESENTATIONS
Proof. Let {πn } be L dense in Rep(Γ, H). Then it is easy to check that {π ∈ Rep(Γ, H) : π ∼ = n πn } is dense in Rep(Γ, H). Indeed fix a nbhd V = {σ : ∀i, j ≤ k∀γ ∈ F (|hσ(γ)(ei ), ej i − hπn0 (γ)(ei ), ej i| < )} of πn0 , where {em } is an L orthonormal basis for H, k = 1, L2, . . . , F ⊆ Γ is finite and > 0. Now π ∈ Rep(Γ, K), where K = n n n Hn , with Hn (n) isomorphic to H. Fix an orthonormal basis {em } for each Hn so that L (n ) T (em 0 ) = em is an isomorphism of Hn0 with H that sends n πn |Hn0 to (n ) H such that S(ei 0 ) = ei for πn0 . Consider then an isomorphism S : K → L i ≤ k, and let σ ∈ Rep(Γ, H) be the image of n πn under S. Then clearly σ ∈V. 2 (B) If π is contained in ρ, i.e., π is isomorphic to a subrepresentation (i.e., restriction to a closed invariant subspace) of ρ, we write π ≤ ρ. We say that π is weakly contained in ρ, in symbols π ≺ ρ, if every positive-definite function realized in π is the pointwise limit of a sequence of finite sums of positive definite functions realized in ρ. Recalling that a positive-definite function realized in a representation σ ∈ Rep(Γ, H) is a function f : Γ → C given by f (γ) = hσ(γ)(v), vi, for some v ∈ H, this means that for any v ∈ Hπ , > 0, F ⊆ Γ finite, there are v1 , . . . , vk ∈ Hρ P such that |hπ(γ)(v), vi − ki=1 hρ(γ)(vi ), vi i| < , ∀γ ∈ F . Clearly ≺ is transitive. We say that π, ρ are weakly equivalent, in symbols, π ∼ ρ, if π ≺ ρ and ρ ≺ π. Clearly π ≤ ρ ⇒ π ≺ ρ, π ∼ = ρ ⇒ π ∼ ρ. Note that π ∼n·π for any n = 1, 2, . . . , ∞, where n · π is the direct sum of n copies of π (and ∞ represents ℵ0 ). One can then show (see, e.g., Kechris [Kec4], 2.2) that π ≺ ρ is equivalent to any of the following: (i) Every positive definite function realized in π is the pointwise limit of a sequence of positive definite functions realized in ∞ · ρ. (ii) For every v1 , . . . , vk ∈ Hπ , > 0, F ⊆ Γ finite, there are w1 , . . . , wk ∈ H∞·ρ such that |hπ(γ)(vi ), vj i − h∞ · ρ(γ)(wi ), wj i| < , ∀γ ∈ F, i, j ≤ k. (iii) Same as (ii) for v1 , . . . , vk orthonormal and requiring that w1 , . . . , wk are also orthonormal. We can use (ii), (iii) to show the following fact. Proposition H.2. Let π ∈ Rep(Γ, Hπ ), ρ ∈ Rep(Γ, Hρ ), where Hπ , Hρ are infinite-dimensional. Then π ≺ ρ ⇔ π ∈ {σ ∈ Rep(Γ, Hπ ) : σ ∼ = ∞ · ρ}.
H. THE SPACE OF UNITARY REPRESENTATIONS
209
Proof. ⇒: Fix an orthonormal basis {en } for Hπ , and a basic nbhd V = {σ : ∀i, j ≤ k∀γ ∈ F |hσ(γ)(ei ), ej i − hπ(γ)(ei ), ej i| < }, (k = 1, 2, . . . , F ⊆ Γ finite, > 0) of π. If π ≺ ρ, then there is an orthonormal set w1 , . . . , wk in H∞·ρ such that |hπ(γ)(ei ), ej i − h∞ · ρ(γ)(wi ), wj i| < , ∀i, j ≤ k, ∀γ ∈ F . Fix then an isomorphism T : Hπ → H∞·ρ with T (ei ) = wi , ∀i ≤ k, and let σ be the copy of ∞ · ρ in Hπ induced by T −1 . Then hσ(γ)(ei ), ej i = h∞ · ρ(γ)(wi ), wj i, so σ ∈ V , thus V ∩ {σ ∈ Rep(Γ, Hπ ) : σ ∼ 6 ∅. = ∞ · ρ} = ⇐: Given v1 , . . . , vk ∈ Hπ , > 0, F ⊆ Γ finite, {σ ∈ Rep(Γ, Hπ ) : ∀i, j ≤ k∀γ ∈ F |hσ(γ)(vi ), vj i − hπ(γ)(vi ), vj i| < } is an open nbhd of π, so it contains σ ∼ = ∞ · ρ. It is clear then that there are w1 , . . . , wk ∈ H∞·ρ with |hπ(γ)(vi ), vj i−h∞· ρ(γ)(wi ), wj i| < , so π ≺ ∞·ρ. 2 Thus π ≺ ρ on Rep(Γ, H), H infinite-dimensional, is, except for replacing ρ by ∞ · ρ, the partial preordering associated with the conjugacy action of U (H) on Rep(Γ, H) (see the paragraph preceding 10.3). Remark. Using the terminology of Bekka-de la Harpe-Valette [BdlHV], F.1.2, one says that π is weakly contained in the sense of Zimmer in ρ, in symbols π ≺Z ρ, if for every v1 , . . . , vn ∈ Hπ , > 0, F ⊆ Γ finite, there are w1 , . . . , wn ∈ Hρ such that |hπ(γ)(vi ), vj i − hρ(γ)(wi ), wj i| < , ∀γ ∈ F, i, j ≤ n. This means that (ii) above holds for ρ instead of ∞ · ρ. It is easy to show (see, e.g., Kechris [Kec4], 2.2) that this is equivalent again to this condition for v1 , . . . , vn orthonormal and requiring that w1 , . . . , wn are also orthonormal. It follows, as in proposition H.2, that for π ∈ Rep(Γ, Hπ ), ρ ∈ Rep(Γ, Hρ ), where Hπ , Hρ are infinite-dimensional, π ≺Z ρ ⇔ π ∈ {σ ∈ Rep(Γ, Hπ ) : σ ∼ = ρ}. Thus ≺Z corresponds exactly to the partial preordering associated with the conjugacy action of U (H) on Rep(Γ, H). We also have that π ≺Z ρ ⇒ π ≺ ρ and π ≺ ρ ⇔ π ≺Z ∞ · ρ. As Bekka-de la Harpe-Valette [BdlHV], F.1.2 point out, if π is irreducible, then π ≺ ρ ⇔ π ≺Z ρ. In particular, 1Γ ≺ π ⇔ 1Γ ≺Z π ⇔ π admits non-0 almost invariant vectors. Corollary H.3. For H infinite-dimensional, the relation π ≺ ρ on the space Rep(Γ, H) is a Gδ subset of Rep(Γ, H)2 . Every section ≺ρ = {π : π ≺ ρ} is closed in Rep(Γ, H). Proof. The first assertion follows from the following easy lemma. Lemma H.4. Let π ∈ Rep(Γ, Hπ ) and ρ ∈ Rep(Γ, Hρ ). Fix a countable dense set D ⊆ Hπ . Then π ≺ ρ iff any positive-definite function of the form γ 7→ hπ(γ)(v), vi, for v ∈ D, is the pointwise limit of a sequence of finite sums of positive-definite functions realized in ρ. The second assertion follows from Proposition H.2.
2
210
H. THE SPACE OF UNITARY REPRESENTATIONS
(C) A representation π ∈ Rep(Γ, H) is irreducible if it admits no nontrivial subrepresentation, i.e., the only invariant closed subspaces are {0}, H. We denote by Irr(Γ, H) the set of irreducible representations in Rep(Γ, H). Proposition H.5. Irr(Γ, H) is Gδ in Rep(Γ, H) (and thus a Polish space in the relative topology). Proof. Let π ∈ Rep(Γ, H). Fix a unit vector h ∈ H and let ϕπ,h (γ) = ϕh (γ) = hπ(γ)(h), hi be the corresponding positive-definite function. Then (see Bekka-de la Harpe-Valette [BdlHV], C.5.2) π is irreducible iff h is a cyclic vector for π and ϕh is an extreme point in the compact convex set P1 (Γ) of all positive-definite functions ϕ : Γ → C with ϕ(1) = 1. Recall that h ∈ H is cyclic if the closed linear span of {π(γ)(h) : γ ∈ Γ} is equal to H. We view here P1 (Γ) as a compact convex subset of the unit ball of `∞ (Γ) = `∞ (Γ, C) with the weak∗ -topology (i.e., pointwise convergence topology). It is clear that the set ext(P1 (Γ)) of extreme points of P1 (Γ) is Gδ in P1 (Γ) and that π 7→ ϕπ,h is continuous from Rep(Γ, H) into P1 (Γ). Thus π ∈ Irr(Γ, H) ⇔ h is cyclic for π and ϕπ,h ∈ ext(P1 (Γ)) is Gδ in Rep(Γ, H).
2
The following result is well-known (see, for example, [Ef2], III or [Hj3]). Theorem H.6. Isomorphism on Irr(Γ, H) is an Fσ equivalence relation (as a subset of Irr(Γ, H)2 ). Proof. We use Schur’s Lemma (see Folland [Fo], 3.5 (b)), which asserts that π1 , π2 ∈ Irr(Γ, H) are isomorphic iff there is a non-0 bounded linear operator S in H such that ∀γ ∈ Γ(Sπ1 (γ) = π2 (γ)S) (such a S is called an intertwining operator for π1 , π2 ). Denote by B1 (H) the set of bounded linear operators on H of norm ≤ 1 with the weak operator topology, i.e., the one induced by the maps T 7→ hT x, yi, for x, y ∈ H. Then B1 (H) is compact metrizable, by Alaoglu’s Theorem and Sunder [Sun], 0.3.1 (b), and U (H) is a Gδ subset of B1 (H). So we have for π1 , π2 ∈ Irr(Γ, H), ∼ π2 ⇔ ∃S ∈ B1 (H)[S 6= 0 and ∀γ ∈ Γ(Sπ1 (γ) = π2 (γ)S)]. π1 = Claim. There is compact K ⊆ B1 (H)3 such that K∩(U (H)2 ×B1 (H)) = {(T1 , T2 , S) : ST1 = T2 S}. Granting this, we have for π1 , π2 ∈ Irr(Γ, H), π1 ∼ = π2 ⇔ ∃S ∈ B1 (H)[S 6= 0 and ∀γ ∈ Γ(π1 (γ), π2 (γ), S) ∈ K)]. Now the topology on Irr(Γ, H) ⊆ Rep(Γ, H) ⊆ U (H)Γ ⊆ B1 (H)Γ is the relative topology from B1 (H)Γ , which is a compact metrizable space, and R = {(π1 , π2 , S) ∈ (B1 (H)Γ )2 × B1 (H) : ∀γ(π1 (γ), π2 (γ), S) ∈ K}
H. THE SPACE OF UNITARY REPRESENTATIONS
211
is compact, so Q = {(π1 , π2 ) ∈ (B1 (H)Γ )2 : ∃S ∈ B1 (H)[S 6= 0 and (π1 , π2 , S) ∈ R]} is Kσ in (B1 (H)Γ )2 , and Q ∩ Irr(Γ, H)2 = (∼ = |Irr(Γ, H)2 ) is Fσ in Irr(Γ, H)2 . Proof of the claim. We need to check that {(T1 , T2 , S) : ST1 = T2 S} is closed in U (H)2 × B1 (H). Notice that ST1 = T2 S ⇔ ∀x, y ∈ H[hST1 (x), yi = hx, S ∗ T2∗ (y)i] and, as U 7→ U ∗ is continuous in B1 (H) (see Sunder [Sun], 0.3.2 (a)), it is enough to check that (S, T ) 7→ hST (x), yi is continuous in B1 (H) × U (H), ∀x, y ∈ H. Assume then that Sn → S, Tn → T weakly, where Sn , S ∈ B1 (H), Tn , T ∈ U (H). Since the weak and strong operator topologies agree on U (H), Tn → T strongly as well, so |hSn Tn (x), yi − hST (x), yi| ≤ |hSn Tn (x), yi − hSn T (x), yi| + |hSn T (x), yi − hST (x), yi| = |hSn (Tn (x) − T (x)), yi| + |hSn (z), yi − hS(z), yi|, where z = T (x), and this is less than or equal to kSn k · kTn (x) − T (x)k · kyk + |hSn (z), yi − hS(z), yi| which converges to 0 as n → ∞, since Tn → T strongly and Sn → S weakly. 2 Remark. Concerning the descriptive complexity of the relation of isomorphism on Rep(Γ, H), it is clearly Borel if Γ is abelian by the Spectral Theorem, but I do not know what is the situation for arbitrary Γ. There are cases where Irr(Γ, H) is a dense Gδ . Proposition H.7 (Hjorth [Hj2], 5.6, for N = ∞). For each N ≥ 2, the set Irr(FN , H) is dense Gδ in Rep(FN , H). Proof. By Section 4, (D), the set of (g1 , . . . , gN ) ∈ U (H)N such that hg1 , . . . , gN i is a dense subgroup of U (H) is dense Gδ in U (H)N . As Rep(FN , H) can be identified with U (H)N , it follows that D = {π ∈ Rep(FN , H) : π(Fn ) is dense in U (H)} is dense Gδ in Rep(FN , H), and clearly D ⊆ Irr(FN , H). 2 On the other hand we will see below, in H.13, that if Γ is infinite and has property (T) and H is infinite-dimensional, then Irr(Γ, H) is nowhere dense.
212
H. THE SPACE OF UNITARY REPRESENTATIONS
Thoma [Th] has shown that if Γ is not abelian-by-finite and H is infinitedimensional, then ∼ = |Irr(Γ, H) is not smooth, and Hjorth [Hj3] extended this by showing that it cannot be classified by countable structures. The arguments in Hjorth [Hj2] in fact show that the action of U (H) on the space Irr(FN , H), N ≥ 2, is generically turbulent. (Thoma [Th] has also shown that when Γ is abelian-by-finite and H is infinite-dimensional, then Irr(Γ, H) = ∅.) We will now give a more detailed version of Hjorth’s result. First let us note the following general fact. Recall that a subset A ⊆ X of a topological space X is called locally closed if it is open in its closure. Proposition H.8. Let a Polish group G act continuously on a Polish space X induced by the action is F (as X and assume the equivalence relation EG σ 2 a subset of X ). Then the set LC = {x ∈ X : G · x is locally closed} is Fσ . S X = Proof. Fix an open basis {Vn } for X. Let EG n Fn , Fn closed in X 2 . For F ⊆ X 2 , x ∈ X, let F (x) = {y : (x, y) ∈ F }. Then x ∈ LC ⇔ G · x is open in G · x [ ⇔ Fn (x) is not meager in G · x n
⇔ ∃n∃m(Vn ∩ G · x 6= ∅ & Vn ∩ G · x ⊆ Fm (x)) ⇔ ∃n∃m(x ∈ G · Vn & ∀g ∈ G(g · x ∈ Vn ⇒ (x, g · x) ∈ Fm )) which clearly shows that LC is Fσ .
2
Consider now the conjugacy action of U (H), H an infinite-dimensional, separable Hilbert space, on Irr(Γ, H). The corresponding equivalence relation is ∼ = on Irr(Γ, H), which, by H.6, is Fσ . Thus the set LC(Γ, H) of irreducible representations with locally closed conjugacy class in Irr(Γ, H) is Fσ in Irr(Γ, H). Effros’ theorem implies that ∼ = |Irr(Γ, H) is smooth iff LC(Γ, H) = Irr(Γ, H), thus, by Thoma [Th], if Γ is not abelian-by-finite the set NLC(Γ, H) = Irr(Γ, H) \ LC(Γ, H) of irreducible representations with non-locally closed conjugacy class in the space Irr(Γ, H) is a non-empty Gδ set and ∼ = |NLC(Γ, H) is also non-smooth ∼ (since = |LC(Γ, H) is easily smooth). I do not know if NLC(Γ, H) is dense in Irr(Γ, H). Finally for each π ∈ Irr(Γ, H), let S(π) = {σ ∈ Irr(Γ, H) : σ ≺ π} = {σ ∈ Irr(Γ, H) : σ ∈ U (H) · π}, the latter equality following from the Remark after H.2. Clearly S(π) is a conjugacy invariant closed subset of Irr(Γ, H). The following is a more detailed version of the main result in Hjorth [Hj3].
H. THE SPACE OF UNITARY REPRESENTATIONS
213
Theorem H.9 (Hjorth [Hj3]). Let Γ be a countable group and assume H is infinite-dimensional. For each π ∈ NLC(Γ, H), the conjugacy action of U (H) on S(π) is generically turbulent. In particular, if Γ is not abelian-byfinite, then ∼ = |Irr(Γ, H) is not classifiable by countable structures. Proof. We will make use of the following standard lemma. Lemma H.10. Let ρ, σ ∈ Irr(Γ, H). Then the following are equivalent: (i) ρ ∼ = σ, (ii) There is a sequence {Tn } in U (H) such that Tn · ρ = Tn ρTn−1 → σ and no subsequence of {Tn } converges in the weak topology of B1 (H) to 0. Proof. (i) ⇒ (ii) is clear, as we can take {Tn } to be a constant sequence. (ii) ⇒ (i): As U (H) ⊆ B1 (H) and the weak topology on B1 (H) is compact metrizable, by going to a subsequence, we can assume that Tn →w T ∈ B1 (H), where T 6= 0. We claim now that T is an intertwining operator for ρ, σ and so, by Schur’s Lemma, ρ ∼ = σ. Since T 7→ T ∗ is continuous in the weak topology, note that we also have Tn∗ = Tn−1 →w T ∗ . We need to show that T ρ(γ −1 ) = σ(γ −1 )T, ∀γ ∈ Γ or equivalently ρ(γ)T ∗ = T ∗ σ(γ), i.e., hρ(γ)T ∗ (x), yi = hT ∗ σ(γ)(x), yi, for γ ∈ Γ, x, y ∈ H. Now as Tn ρ(γ)Tn−1 → σ(γ) in the topology of U (H), clearly kTn ρ(γ)Tn−1 (x) − σ(γ)(x)k → 0, so kρ(γ)Tn−1 (x) − Tn−1 σ(γ)(x)k → 0, thus |hρ(γ)Tn−1 (x), yi − hTn−1 σ(γ)(x), yi| → 0. But hρ(γ)Tn−1 (x), yi = hTn−1 (x), ρ(γ)−1 (y)i → hT ∗ (x), ρ(γ)−1 (y)i = hρ(γ)T ∗ (x), yi and hTn−1 σ(γ)(x), yi → hT ∗ σ(γ)(x), yi, so we are done.
2
We thus have that if σ ∈ S(ρ) (i.e., σ ∈ U (H) · ρ) and ρ ∼ 6 σ, then there = is {Tn } in U (H) with Tn · ρ → σ and Tn →w 0. We proceed to verify that the conjugacy action of U (H) on S(π) is generically turbulent for the given π ∈ NLC(Γ, H). First it is clear that the conjugacy class of π is dense and meager in S(π) (else it would be open in S(π) contradicting that π ∈ NLC(Γ, H)). If ρ ∈ S(π) \ U (H) · π, then again U (H) · ρ is meager in S(π) else it would have to be open and therefore intersect U (H)·π. Thus we only need to verify that π is a turbulent point for the conjugacy action of U (H) on S(π). Keep in mind below that S(π) \ U (H) · π is dense in S(π).
214
H. THE SPACE OF UNITARY REPRESENTATIONS
Claim. For any non-empty open set W in S(π) and any orthonormal set e1 , . . . , ep in H, there is an orthonormal set e1 , . . . , ep , ep+1 , . . . , eq extending e1 , . . . , ep , and T ∈ U (H) such that (a) T (ei ) ⊥ ej , ∀i, j ≤ q, (b) T 2 (ei ) = −ei , ∀i ≤ q, (c) T = id on (H0 ⊕ T (H0 ))⊥ , where H0 = he1 , . . . , eq i, (d) T · π ∈ W . Assume this and proceed to show that π is turbulent. Fix a basic nbhd of π in S(π): W = Wπ,e,F, = {ρ : ∀γ ∈ F ∀i ≤ k|hρ(γ)(ei ), ej i − hπ(γ)(ei ), ej i| < }, where e = e1 , . . . , ek is orthonormal, F ∈ Γ is finite symmetric, and > 0. Let e1 , . . . , ek , ek+1 , . . . , ep be an orthonormal basis for the span of e1 , . . . , ek , π(γ)(ei ), 1 ≤ i ≤ k, γ ∈ F, and let e1 , . . . , ep , ep+1 , . . . , eq , T be as in the claim (so that in particular T · π ∈ W ). It is clearly enough to find a continuous path {Tθ }0≤θ≤π/2 in U (H) with T0 = 1, Tπ/2 = T and Tθ · π ∈ W, ∀θ (which clearly implies turbulence as in the proof of 5.1). Put Tθ (ei ) = (cos θ)ei + (sin θ)T (ei ), Tθ (T (ei )) = (− sin θ)ei + (cos θ)T (ei ), for i = 1, . . . , q, and Tθ = id on (H0 ⊕ T (H0 ))⊥ , where H0 = he1 , . . . , eq i. Thus T0 = 1, Tπ/2 = T and we will verify that Tθ · π ∈ W . Indeed for i, j ≤ k, γ ∈ F, hTθ · π(γ)(ei ), ej i = hπ(γ)(Tθ−1 (ei )), Tθ−1 (ej )i = hπ(γ)((cos θ)ei − (sin θ)T (ei )), (cos θ)ej − (sin θ)T (ej )i = (cos2 θ)hπ(γ)(ei ), ej i + (sin2 θ)hπ(γ)(T (ei )), T (ej )i, since π(γ)(T (ei )) ⊥ ej and π(γ)(ei ) ⊥ T (ej ) (note that π(γ)(T (ei )) ⊥ ej iff T (ei ) ⊥ π(γ)−1 (ej ) = π(γ −1 )(ej ), and γ −1 ∈ F ). So if A = hTθ · π(γ)(ei ), ej i − hπ(γ)(ei ), ej i, then A = − sin2 θ(hπ(γ)(ei ), ej i − hπ(γ)(T (ei )), T (ej )i) = − sin2 θ(hπ(γ)(ei ), ej i − hπ(γ)(−T −1 (ei )), −T −1 (ej )i) = − sin2 θ(hπ(γ)(ei ), ej i − hπ(γ)(T −1 (ei )), T −1 (ej )i) = − sin2 θ(hπ(γ)(ei ), ej )i − hT π(γ)T −1 (ei ), ej i), therefore |A| ≤ |hT · π(γ)(ei ), ej i − hπ(γ)(ei ), ej i| < , so Tθ · π ∈ W .
H. THE SPACE OF UNITARY REPRESENTATIONS
215
Proof of the Claim. Without loss of generality we can assume that W has the form W ≡ Wσ,e,F, (where Wσ,e,F, is defined as above) for some e = e1 , . . . , ep , ep+1 , . . . , eq , and some σ ∈ S(π) \ U (H) · π (as S(π) \ U (H) · π is dense in S(π)). Thus there is Tn ∈ U (H) with Tn · π → σ and Tn →w 0, thus also Tn−1 →w 0. Note that since hTn−1 (ei ), ej i → 0, ∀i, j ≤ q, for all large enough n, e1 , . . . , eq , Tn−1 (−e1 ), . . . , Tn−1 (−eq ) are linearly independent. Apply Gram-Schmid to e1 , . . . , eq , Tn−1 (−e1 ), . . . , Tn−1 (−eq ) to get an orthonor(n) (n) mal set e1 , . . . , eq , f1 , . . . , fq . Since hTn−1 (−ei ), ej i → 0, ∀i, j ≤ q, it fol(n) lows that kTn−1 (−ei ) − fi k → 0, ∀i ≤ q. Define Sn ∈ U (H) by (n)
Sn (ei ) = fi , (n) Sn (fi ) = −ei , ∀i ≤ q, and (n)
Sn = id on he1 , . . . , eq , f1 , . . . , fq(n) i⊥ . (n)
(n)
Then Sn (ei ) = fi ⊥ ej , Sn2 (ei ) = Sn (fi ) = −ei , ∀i, j ≤ q and kSn−1 (ei ) − (n) Tn−1 (ei )k = kTn−1 (−ei ) − Sn−1 (−ei )k = kTn−1 (−ei ) − fi k → 0, ∀i ≤ q. Thus for all large enough n, kSn−1 (ei ) − Tn−1 (ei )k < /4, ∀i ≤ q, and Tn · π ∈ Wσ,e,F,/2 , i.e., |hTn π(γ)Tn−1 (ei ), ej i − hσ(γ)(ei ), ej i| < /2, ∀i, j ≤ q, ∀γ ∈ F , or equivalently |hπ(γ)(Tn−1 (ei )), Tn−1 (ej )i − hσ(γ)(ei ), ej i| < /2. Thus |hπ(γ)(Sn−1 (ei )), Sn−1 (ej )i − hσ(γ)(ei ), ej i| < , ∀i, j ≤ q, ∀γ ∈ F , i.e., Sn · π ∈ W, so Sn = T works.
2
It follows that for infinite-dimensional H the conjugacy action of U (H) on Irr(Γ, H) is generically turbulent iff there is π ∈ NLC(Γ, H) with S(π) = Irr(Γ, H), i.e., there is π ∈ NLC(Γ, H) with dense conjugacy class in Irr(Γ, H) or equivalently ∀ρ ∈ Irr(Γ, H)(ρ ≺ π). There are countable groups Γ for which there is no dense conjugacy class in Irr(Γ, H), e.g., products Γ = G × ∆, where ∆ is non-trivial abelian and G is not abelian-by-finite. ˆ Indeed for such a Γ and π ∈ Irr(G, H), H infinite-dimensional, and χ ∈ ∆ (i.e., χ a character on ∆), define πχ ∈ Irr(Γ, H), by πχ (g, δ)(h) = χ(δ)π(g)(h). Clearly the map ˆ → Irr(Γ, H) Φ : Irr(G, H) × ∆
216
H. THE SPACE OF UNITARY REPRESENTATIONS
is 1–1 and its range is conjugacy invariant, as T · πχ = (T · π)χ for π ∈ Irr(G, H), T ∈ U (H). Moreover it is well known (since an abelian group is of type I) that every ρ ∈ Irr(Γ, H) is isomorphic to some πχ (see Folland [Fo], 7.25), so Φ is a bijection and it is easy to check that it is actually a ˆ with Irr(Γ, H) which preserves conjugacy. homeomorphism of Irr(G, H) × ∆ ˆ (on So identifying πχ with (π, χ), we can identify Irr(Γ, H) with Irr(G, H)×∆ which U (H) acts by conjugacy on the first coordinate). Clearly S(π, χ) = S(π) × {χ}, so no conjugacy class is dense. In the opposite direction, if Irr(Γ, H) is dense in Rep(Γ, H), then there is π ∈ Irr(Γ, H) with dense conjugacy class in Rep(Γ, H) and thus in Irr(Γ, H). (This is because both Irr(Γ, H) and {π ∈ Rep(Γ, H) : π has a dense conjugacy class} are dense Gδ in Rep(Γ, H).) So if moreover every conjugacy class in the space Rep(Γ, H) is meager, the conjugacy action of U (H) in Irr(Γ, H) is generically turbulent. This is for example the case when Γ = FN , N ≥ 2, by H.7. (Note that Rep(FN , H) can be identified with U (H)N and so the conjugacy classes in Rep(FN , H) are meager, by the remark following 2.5.) The fact that when Γ = FN , N ≥ 2, there is π ∈ Irr(Γ, H) with dense conjugacy class in Rep(Γ, H), means that there is π ∈ Irr(Γ, H) so that ρ ≺Z π, ∀ρ ∈ Rep(Γ, H), and in particular implies that there is an irreducible representation weakly containing any representation of Γ, a result first proved by Yoshizawa [Y]. More generally, call a countable group Γ densely representable (DR) if the set of all π ∈ Rep(Γ, H) with range π(Γ) = {π(γ) : γ ∈ Γ} ⊆ U (H) dense in U (H) is dense (and thus dense Gδ ) in Rep(Γ, H). Since every π ∈ Rep(Γ, H) with dense range is clearly irreducible, it follows that if Γ is DR, then Irr(Γ, H) is dense Gδ in Rep(Γ, H). All free groups FN , N ≥ 2, are DR (see the second paragraph after 4.12). However, as we have seen earlier, no product group G × ∆, with ∆ non-trivial abelian and G not abelian-byfinite, is DR. I do not know if, for example, F2 × F2 is DR, in fact, apart from some small variations of free non-abelian groups, I do not know any other examples of DR groups. In conclusion it is unclear what groups Γ admit a dense conjugacy class in Irr(Γ, H), what groups are DR and finally what groups Γ have the property that conjugacy classes in Rep(Γ, H) are meager. Remark. Let Γ be a non-amenable group and denote by λΓ its leftregular representation on `2 (Γ). Note that (in fact for any non-trivial Γ) λΓ is not irreducible; see de la Harpe [dlH], Appendix II. We consider the space S(λΓ ) of all irreducible representations (on an infinite-dimensional H) weakly contained in λΓ . Clearly this is a closed subspace of Irr(Γ, H). There is a dense conjugacy class in S(λΓ ) iff λΓ is weakly equivalent to an irreducible representation. It is known, see de la Harpe [dlH], 21. Proposition, that if Γ is ICC (i.e., Γ has infinite conjugacy classes), then λΓ is weakly
H. THE SPACE OF UNITARY REPRESENTATIONS
217
equivalent to an irreducible representation and therefore S(λΓ ) has a dense conjugacy class. If ∼ = |S(λΓ ) is not smooth, then NLC(Γ, H) ∩ S(λΓ ) 6= ∅, and thus, by H.9, S(λΓ ) is not classifiable by countable structures. Finally, there are many groups Γ for which the conjugacy action of U (H) on S(λΓ ) is actually minimal (every orbit is dense). This means that any π ∈ S(λΓ ) is weakly equivalent to λΓ . Such groups are called C ∗ -simple since this condition is equivalent to the simplicity of the reduced C ∗ -algebra of Γ. See de la Harpe [dlH] for an extensive discussion of these concepts. (See also the recent preprint Poznansky [Poz], where a characterization of the linear C ∗ -simple groups is obtained.) Among the C ∗ -simple groups are the free groups FN , N ≥ 2. Now if Γ is a C ∗ -simple group, every conjugacy class in S(λΓ ) is meager in S(λΓ ), otherwise it would have to be open in S(λΓ ), so S(λΓ ) would consist of a single conjugacy class, a contradiction. This means that the conjugacy action of U (H) on S(λΓ ) is turbulent (and not just generically turbulent). Remark. As in Section 15, we can also derive some consequences concerning connectedness. Let Γ be a countable group and H be infinitedimensional. Let π ∈ NLC(Γ, H). Then if C ⊆ S(π) contains the conjugacy class of π, C is path connected and locally path connected. Thus if there is π ∈ NLC(Γ, H) with dense conjugacy class in Irr(Γ, H), as in the case Γ = FN , N ≥ 2, then Irr(Γ, H) is path connected and locally path connected. On the other hand, if Γ = G × ∆, ∆ a non-trivial abelian group ˆ is not with an element of finite order and G not abelian-by-finite, then ∆ ∼ ˆ connected (see Rudin [Ru2], 2.5.6), so Irr(Γ, H) = Irr(G, H) × ∆ is not connected. I do now know what groups Γ have the property that Irr(Γ, H) is (path) connected. (D) A representation π ∈ Rep(Γ, H), for infinite Γ, is called ergodic if it has no non-0 invariant vectors, weak mixing if it contains no non-0 finite dimensional subrepresentation, mild mixing if for every h ∈ H, h 6= 0, γn → ∞, we have π(γn )(h) 6→ h, and mixing if it is a c0 -representation, i.e., limγ→∞ hπ(γ)(h), hi = 0, for every h ∈ H. Denote by ERG(Γ, H), MMIX(Γ, H), WMIX(Γ, H), MIX(Γ, H) the corresponding classes of representations. (We interpret these as being empty if Γ is finite.) Clearly ERG(Γ, H) ⊇ WMIX(Γ, H) ⊇ MMIX(Γ, H) ⊇ MIX(Γ, H). See Glasner [Gl2], Ch. 3 and Bergelson-Rosenblatt [BR] for other characterizations of these classes. Using an argument similar to that of the (alternative) proof of 12.1, and Glasner [Gl2] or Bergelson-Rosenblatt [BR], we have the following fact. Proposition H.11. The sets ERG(Γ, H) and WMIX(Γ, H) are Gδ in the space Rep(Γ, H).
218
H. THE SPACE OF UNITARY REPRESENTATIONS
Clearly MIX(Γ, H) is Π03 and MMIX(Γ, H) is Π11 . However, by the paragraph before the Remark in Section 10, (G), for infinite-dimensional H, MMIX(Γ, H) is not Borel, if Γ has an element of infinite order. Also ERG(Γ, H) has no interior in Rep(Γ, H), if H is infinite-dimensional. (E) We now note some analogs of the characterizations in Sections 11, 12. Below 1Γ is the trivial 1-dimensional representation of Γ. The equivalence of (a), (d), (e), (f), (g) in H.12 (i) below was also independently proved in Kerr-Pichot [KP]. Theorem H.12. (i) The following are equivalent for any infinite countable group Γ and infinite-dimensional H: (a) Γ has property (T), (b) ERG(Γ, H) = {π ∈ Rep(Γ, H) : 1Γ 6≺ π}, (c) WMIX(Γ, H) ⊆ {π ∈ Rep(Γ, H) : 1Γ 6≺ π}, (d) ERG(Γ, H) is closed in Rep(Γ, H), (e) ERG(Γ, H) is not dense in Rep(Γ, H), (f ) WMIX(Γ, H) is closed in Rep(Γ, H), (g) WMIX(Γ, H) is not dense in Rep(Γ, H). (ii) The following are equivalent for any infinite countable group Γ and infinite-dimensional H: (a) Γ does not have (HAP). (b) MIX(Γ, H) ⊆ {π ∈ Rep(Γ, H) : 1Γ 6≺ π}. (iii) (Bergelson-Rosenblatt [BR]) If an infinite countable group Γ has (HAP), then the set MIX(Γ, H) is dense in Rep(Γ, H). Proof. (i): (a) ⇔ (b) follows from the definition of property (T). (b) ⇒ (c) is clear and for ¬(b) ⇒ ¬(c) see the remark following 11.2. So (a) ⇔ (b) ⇔ (c). (a) ⇒ (d) follows exactly as in the proof of 12.2, (i) and (d) ⇒ (e) ⇒ (g) is clear. To prove that (d) ⇒ (f), recall that for any π1 ∈ Rep(Γ, H), π2 ∈ Rep(Γ, H2 ), the tensor product π1 ⊗ π2 ∈ Rep(Γ, H1 ⊗ H2 ) is defined by (π1 ⊗ π2 )(γ)(v1 ⊗ v2 ) = π1 (γ)(v1 ) ⊗ π2 (γ)(v2 ). Then π ∈ Rep(Γ, H) is weak mixing iff for every ρ ∈ Rep(Γ, H 0 ) the tensor product π ⊗ ρ is ergodic (see Glasner [Gl2], 3.5). Since (π, ρ) 7→ π ⊗ ρ is continuous, it is clear that (d) ⇒ (f). Again (f) ⇒ (g) is obvious. Finally to prove (g) ⇒ (a), assume that Γ does not have property (T) in order to show that WMIX(Γ, H) = Rep(Γ, H). Fix σ ∈ Rep(Γ, H). By (a) ⇔ (c) there is π0 ∈ Rep(Γ, H) with 1Γ ≺ π0 and π0 ∈ WMIX(Γ, H). Then π0 ⊗ σ is weak mixing. Since 1Γ ≺ π0 , we have 1Γ ⊗ σ ∼ = σ ≺ π0 ⊗ σ (see Bekka-de la HarpeValette [BdlHV], F.3.2), so, by H.2, σ is in the closure of the set of isomorphic copies of ∞ · (π0 ⊗ σ) in Rep(Γ, H). Since π0 ⊗ σ is weak mixing, it is clear that ∞ · (π0 ⊗ σ) is weak mixing, thus σ ∈ WMIX(Γ, H), i.e., Rep(Γ, H) = WMIX(Γ, H).
H. THE SPACE OF UNITARY REPRESENTATIONS
219
(ii): (a) ⇔ (b) is one of the equivalent definitions of (HAP) (see Cherix et al. [CCJJV]). (iii): Assume now that Γ has (HAP). Then by (ii), (b) there is mixing π ∈ Rep(Γ, H) with 1Γ ≺ π. Then for any σ ∈ Rep(Γ, H), σ ∼ = 1Γ ⊗σ ≺ π⊗σ and π ⊗ σ, ∞ · (π ⊗ σ) are mixing, thus σ ∈ MIX(Γ, H). 2 Corollary H.13. If an infinite countable group Γ has property (T) and H is infinite-dimensional, then Irr(Γ, H) is nowhere dense. Proof. Clearly Irr(Γ, H) ⊆ ERG(Γ, H), as H is infinite-dimensional. 2 Remark. Bekka [B] has introduced the concept of an amenable unitary representation. One of its equivalent formulations is the following: π ∈ Rep(Γ, H) is amenable iff 1Γ ≺ π ⊗ π. Here π ∈ Rep(Γ, H) is the conjugate representation defined as follows: H is the conjugate space of H, which can be viewed as H with the same addition but with scalar multiplication and inner product defined by: (λ, x) 7→ λx, hx, yiH = hy, xiH , ∀λ ∈ C, ∀x, y ∈ H. Finally π = π set theoretically. Denote by AMEN(Γ, H) the set of amenable representations of Γ on H. Since π 6∈ WMIX(Γ, H) iff 1Γ ≤ π ⊗ π (see, e.g., Bekka-de la HarpeValette [BdlHV], A.1.11), it follows that if ∼ AMEN(Γ, H) = Rep(Γ, H) \ AMEN(Γ, H), then ∼ AMEN(Γ, H) ⊆ WMIX(Γ, H). Bekka [B], 5.9 shows that if Γ has property (T), then ∼ AMEN(Γ, H) = WMIX(Γ, H) and BekkaValette [BV] show that conversely ∼ AMEN(Γ, H) = WMIX(Γ, H), implies that Γ has property (T). We note that this also follows immediately from H.12. Indeed if Γ does not have property (T), then, by H.12, (i), there is π ∈ WMIX(Γ, H) with 1Γ ≺ π, so that π is amenable. We have seen that Γ has property (T) iff WMIX(Γ, H) is closed. I do not know if it is true that Γ has property (T) iff ∼ AMEN(Γ, H) is closed. Remark. A countable group Γ is said to have property (FD) (see Lubotzky-Shalom [LS]) if the unitary representations which factor through a finite quotient of Γ are dense in Rep(Γ, H). It was shown in LubotzkyShalom [LS] that each free group Fn has property (FD). We note that this follows from the fact that Aut(X, µ) is topologically locally finite together with Appendix E (the latter being also used in [LS]). Observe that the property (FD) for Fn is equivalent to the following property of U (H): The set of (g1 , . . . , gn ) ∈ U (H)n such that hg1 , . . . , gn i is finite is dense in U (H)n . To prove this, fix g1 , . . . , gn ∈ U (H)n , e1 , . . . , em ∈ H and > 0. We will find (h1 , . . . , hn ) ∈ U (H)n with hh1 , . . . , hn i finite and kgi (ej ) − hi (ej )k < , i = 1, . . . , n, j = 1, . . . , m. By Appendix E, there is a standard Borel space (X, µ) such that H is a closed subspace of L20 (X, µ) = H0 and g1 , . . . , gn extend to g¯1 , . . . , g¯n ∈ Aut(X, µ) ⊆ U (L20 (X, µ)) = U (H0 ). Use now the topological local finiteness of Aut(X, µ) to find g˜1 , . . . , g˜n ∈ U (H0 ) with kgi (ej ) − g˜i (ej )k < , ∀i ≤ n, ∀j ≤ m and hg˜1 , . . . , g˜n i finite. We can find an isomorphism π : H0 → H with π(ej ) = ej , π(gi (ej )) = gi (ej ), ∀i ≤ n, ∀j ≤ m. Finally let hi ∈ U (H) be the image of g˜i under π.
220
H. THE SPACE OF UNITARY REPRESENTATIONS
This clearly works as hi (ej ) = π(g˜i (ej )), gi (ej ) = π(gi (ej )) and so khi (ej ) − gi (ej )k = kπ(g˜i (ej )) − π(gi (ej ))k = kg˜i (ej ) − gi (ej )k < . In connection with this argument, I do not know if U (H) itself is topologically locally finite. (F) Concerning the connection between Rep(Γ, H) and A(Γ, X, µ), let us call π ∈ Rep(Γ, H) (H infinite-dimensional) realizable by an action if there is a ∈ A(Γ, X, µ) such that π ∼ = κa0 (= the Koopman representation on 2 L0 (X, µ) associated with a). Proposition H.14. Let Γ be a countable group and assume that H is infinite-dimensional. Then the realizable by an action representations are dense in Rep(Γ, H). Proof. Let Γ = {γi }∞ i=1 . Fix π ∈ Rep(Γ, H) and an orthonormal basis {ej }∞ for H. Given > 0, n ≥ 1, m ≥ 1, it is enough to find a realizable by j=1 an action ρ ∈ Rep(Γ, H) such that kπ(γi )(ej )−ρ(γi )(ej )k < , ∀i ≤ n, j ≤ m. By E.1 we can assume that H is a closed subspace of L20 (X, µ) and there is a ∈ A(Γ, X, µ) with κa0 extending π. Let H1 be a finite-dimensional subspace of H which contains all ej , π(γi )(ej ), i ≤ n, j ≤ m. Then find an isomorphism T : L20 (X, µ) → H such that T |H1 = identity. Let ρ ∈ Rep(Γ, H) be the image of κa0 under T , so that clearly ρ is realizable by an action. We have for i ≤ n, j ≤ m, ρ(γi )(ej ) = T (κa0 (γi )(ej )) = T (π(γi )(ej )) = π(γi )(ej ), so we are done.
2
Corollary H.15. Let Γ be a countable group and H be infinite-dimensional. Let P ⊆ Rep(Γ, H) be invariant under conjugacy and assume that {a ∈ A(Γ, X, µ) : ∃π ∈ P (π ∼ = κa0 )} is dense in (A(Γ, X, µ), w). Then P is dense in Rep(Γ, H). Proof. We can assume that H = L20 (X, µ), in which case (A(Γ, X, µ), w) is a closed subspace of Rep(Γ, H) (via the identification of a ∈ A(Γ, X, µ) with κa0 ). Then H.14 shows that the saturation of A(Γ, X, µ) under the conjugacy action of U (H) is dense in Rep(Γ, H). Our hypothesis implies that P ∩ A(Γ, X, µ) is dense in (A(Γ, X, µ), w) therefore, as the conjugacy action is continuous, P is dense in Rep(Γ, H). 2 Since the set of realizable by an action π ∈ Rep(Γ, H) is invariant under conjugacy, it is either meager or comeager in Rep(Γ, H). When Γ is torsionfree abelian, it is not hard to see that it is meager. Indeed if π is realizable by an action, then the maximal spectral type σ of π (see Appendix F) is symmetric up to measure equivalence, σ ∼ σ −1 , where σ −1 (A) = σ(A−1 ), ˆ (see, e.g., Nadkarni [Na], p. 21 or Lema´ for A a Borel subset of Γ nczyk [Le], pp. 83–84, for the Γ = Z case). On the other hand, the set of π ∈
H. THE SPACE OF UNITARY REPRESENTATIONS
221
Rep(Γ, H) whose maximal spectral type σ satisfies σ ⊥ σ −1 is Gδ , since there is continuous π 7→ σπ with σπ the maximal spectral type of π (see Appendix ˆ is a Gδ relation (note that µ ⊥ ν iff µ ∧ ν = 0 iff F, (B)), andR ⊥ on P (Γ) R ˆ 0 ≤ f ≤ 1} = 0). Now one can also see inf{ f dµ + (1 − f )dν : f ∈ C(Γ), that the set of all π ∈ Rep(Γ, H) whose maximal spectral type σ satisfies σ ⊥ σ −1 is dense in Rep(Γ, H), which completes the proof. To verify this, ˆ consider the representation πµ of Γ on L2 (Γ, ˆ µ) defined for each µ ∈ P (Γ) ˆ by πµ (γ)(f ) = γ¯ f , where γ ∈ Γ is viewed as a character of Γ (see Appendix F). Clearly µ is the maximal Lspectral type of πµ . Now any π ∈ Rep(Γ, H) is isomorphic to a direct sum n πµn (see Appendix F), and by approximating ˆ by probability measures with finite support, we probability measures on Γ ˆ there is ν1 , ν2 , . . . , ∈ P (Γ) ˆ see that for each µ1 , µ2 , · · · ∈ P (Γ), that P −n P −n such −1 eachL νn has finite 2 νn ⊥ ( 2 νL n ) . Thus L support, πµn ≺ πνn , and π . Moreover the maximal spectral type of π∼ = n πµ n ≺ ν n n πνn and P L n thus of ∞ · ( n πνn ) is σ = n 2−n νn and, as σ ⊥ σ −1 , it follows from H.2 that π is in the closure of the set of all ρ ∈ Rep(Γ, H) whose maximal spectral type σ satisfies σ ⊥ σ −1 . In the case Γ = Z, one can also give a proof along the lines of the proof of 2.5 and the remark following it. We take H = L20 (X, µ) and we identify Rep(Z, H) with U (L20 (X, µ)). Then the set of representable by an action U ∈ U (L20 (X, µ)) is the saturation under conjugacy of Aut(X, µ), which we view as a closed subgroup of U (L20 (X, µ)). Assuming this set is comeager, towards a contradiction, we can find a Borel function U 7→ fU on U (H) such that ∀∗ U (fU U fU−1 ∈ Aut(X, µ)). Fix now a real-valued ϕ ∈ H with kϕk = 1 and {ξn } ⊆ H dense. Then there is some n and open nonempty W ⊆ U (H) such that, letting ξ = ξn , we have 1 −1 ∗ ∀ U ∈ W kfU (ϕ) − ξk < , 8 so that, in particular, 87 < kξk < 89 . Let now 9 n Ω = U ∈ U (H) : ∃n Im hU (ξ), ξi > . 16 This is clearly open and we claim that is also dense. Using the notation of the Remark following 2.5, it is enough to show that it is dense in each U (m) and this follows as in the argument in that remark by approximating elements of U (m) by nth roots of the operator iI. It follows that there is U ∈ U (H) such that kfU−1 (ϕ) − ξk < 81 , V = 9 fU U fU−1 ∈ Aut(X, µ) and Im hU n (ξ), ξi > 16 , for some n. Thus we have Im hV n (fU (ξ)), fU (ξ)i >
9 16
and, since kfU (ξ) − ϕk < 18 , it follows that |hV n (fU (ξ)), fU (ξ)i − hV n (ϕ), ϕi| <
1 8
· kξk +
1 8
· kϕk <
1 8
·
9 8
+
1 8
< 13 ,
222
H. THE SPACE OF UNITARY REPRESENTATIONS
9 therefore Im hV n (ϕ), ϕi > 16 − 31 > 0, contradicting the fact that the number n hV (ϕ), ϕi is real, as ϕ is real and V ∈ Aut(X, µ).
Problem H.16. Let Γ be an infinite countable group and H be infinitedimensional. Is the set of realizable by an action π ∈ Rep(Γ, H) meager in Rep(Γ, H)?
APPENDIX I
Semidirect products of groups (A) Let A, G be groups and let A act on G by automorphisms. Denote the action by a·g. Then the semidirect product AnG is the group defined as follows: The space of A n G is the product A × G. Multiplication is defined by (1)
(a1 , g1 )(a2 , g2 ) = (a1 a2 , g1 (a1 · g2 ))
There is a canonical action of A by automorphisms on AnG that extends the action of A on G, namely a · (b, g) = (aba−1 , a · g) An affine automorphism of G is any map ψ of the form ψ(g) = hϕ(g), where h ∈ G and ϕ is an automorphism of G. Clearly any such ψ can be identified with the pair (ϕ, h), ϕ being the automorphic part of ψ and h = ψ(1) the translation part. Note that if ψ1 = (ϕ1 , h1 ), ψ2 = (ϕ2 , h2 ), then ψ1 (ψ2 (g)) = h1 ϕ1 (ψ2 (g)) = h1 ϕ1 (h2 ϕ2 (g)) = h1 ϕ1 (h2 )ϕ1 ϕ2 (g), i.e., ψ1 ψ2 = (ϕ1 ϕ2 , h1 ϕ1 (h2 )), thus the group of affine automorphisms of G can be identified with the semidirect product of the automorphism group of G with G, where the automorphism group of G acts by evaluation on G. If now A acts by automorphisms on G, then A n G acts by affine automorphisms on G via (a, g) · h = g(a · h), and if A acts on G faithfully (i.e., a · g = g, ∀g ⇒ a = 1), so that A can be viewed as a group of automorphisms of G, then A n G acts faithfully on G, so A n G can be viewed as a group of affine automorphisms of G. Example. Let H be a separable real Hilbert space, viewed as an additive group, A = O(H) its orthogonal group and consider the action of O(H) on H by evaluation. Every isometry ϕ of H can be uniquely written as ϕ = T + b, where T ∈ O(H) and b = ϕ(0) ∈ H. Thus identifying ϕ with the pair (T, b), we see that the isometry group of H is identified with O(H) n H. 223
224
I. SEMIDIRECT PRODUCTS OF GROUPS
Example. Let G = (MALGµ , ∆) and consider the metric dµ (A, B) = µ(A∆B) on G. Then A = Aut(X, µ) acts by isometric automorphisms on G. Every isometry of (MALGµ , dµ ) can be uniquely written as ϕ = T ∆B, where T ∈ Aut(X, µ), B = ϕ(∅) ∈ MALGµ (see Section 1, (B)). Thus the isometry group of (MALGµ , dµ ) is identified with A n G. Denote now by Hom(Γ, K) the set of homomorphisms of a group Γ into a group K. Any ψ ∈ Hom(Γ, A n G) can be viewed as a pair ψ = (ϕ, α), where ψ(γ) = (ϕ(γ), α(γ)). Then ϕ ∈ Hom(Γ, A) and α : Γ → G satisfies α(γδ) = α(γ)(ϕ(γ) · α(δ)). Such an α is called a cocycle associated to the homomorphism ϕ. We denote by Z 1 (ϕ) the set of these cocycles. Thus Hom(Γ, A n G) = {(ϕ, α) : ϕ ∈ Hom(Γ, A) & α ∈ Z 1 (ϕ)}. For each fixed ϕ ∈ Hom(Γ, A), Z 1 (ϕ) gives all possible “extensions” of ϕ to a homomorphism into A n G. We can view ϕ ∈ Hom(Γ, A) as a representation of Γ as a group of automorphisms of G and then Z 1 (ϕ) gives representations of Γ as a group of affine automorphisms of G with automorphic part equal to ϕ. If for instance A = Aut(G) is the automorphism group of G acting by evaluation on G, then Z 1 (ϕ) gives all possible representations of Γ as a group of affine automorphisms of G with automorphic part equal to ϕ. A coboundary of ϕ ∈ Hom(Γ, A) is a cocycle of the form α(γ) = g −1 (ϕ(γ) · g), for some g ∈ G. Their set is denoted by B 1 (ϕ). The group G acts on Z 1 (ϕ) by (g · α)(γ) = gα(γ)(ϕ(γ) · g −1 ) Two cocycles α, β ∈ Z 1 (ϕ) are called cohomologous, in symbols α ∼ β, if ∃g ∈ G(g · α = β). Thus the coboundaries are the cocycles cohomologous to the trivial cocycle 1. The quotient space H 1 (ϕ) = Z 1 (ϕ)/ ∼ is called the (1st)-cohomology space of ϕ. When G is abelian, Z 1 (ϕ) is an abelian group under pointwise multiplication, B 1 (ϕ) is a subgroup and H 1 (ϕ) = Z 1 (ϕ)/B 1 (ϕ) is the (1st)-cohomology group of ϕ. Note here the following simple fact. Proposition I.1. α ∈ Z 1 (ϕ) is a coboundary iff the affine representation of Γ in G, ψ = (ϕ, α), associated to ϕ, α, has a fixed point. Proof. If ψ has a fixed point g −1 ∈ G, then for every γ, ψ(γ)(g −1 ) = g −1 , so α(γ)(ϕ(γ) · g −1 ) = g −1 , thus α(γ) = g −1 (ϕ(γ) · g −1 )−1 = g −1 (ϕ(γ) · g),
I. SEMIDIRECT PRODUCTS OF GROUPS
225
so α ∈ B 1 (ϕ). Conversely, if α ∈ B 1 (ϕ), then α(γ) = g −1 (ϕ(γ) · g) for some g and, reversing the above steps, g −1 is a fixed point of ψ. 2 Example. Let H be a separable real Hilbert space. Then O(H) n H is its isometry group. A homomorphism π of Γ into O(H) is an orthogonal representation of Γ in H. Then Z 1 (π) consists of all the cocycles b : Γ → H (i.e., maps satisfying b(γδ) = b(γ) + π(γ)(b(δ))) and the pairs (π, b) with b ∈ Z 1 (π) give all affine isometric representations of Γ in H with orthogonal part π. Coboundaries are the cocycles of the form b(γ) = π(γ)(h)−h for some h ∈ H. Clearly Z 1 (π) is an abelian group (under pointwise addition), in fact a real vector space, and B 1 (π) a subspace. Then H 1 (π) = Z 1 (π)/B 1 (π) is the (1st-)cohomology group (or vector space) of the representation π. It vanishes (i.e., H 1 (π) = {0}) iff all isometric extensions of π admit fixed points. For more on the cohomology of representations, especially in connection with property (T), see Bekka-de la Harpe-Valette [BdlHV]. Remark. When A, G are Polish groups and A acts continuously on G, then the semidirect product A n G with the product topology is clearly a Polish group. (B) Sometimes one uses a different rule of multiplication for the semidirect product A n G, namely (2)
(a1 , g1 )(a2 , g2 ) = (a1 a2 , (a−1 2 · g1 )g2 ).
It is easy to see however that (a, g) 7→ (a, a−1 · g) is an isomorphism between A n G using the multiplication (1) with A n G using the multiplication (2). It is more convenient to use the second form in the context of the theory of cocycles for group actions. When we use (2) as the multiplication in A n G, it is easy to see that the canonical action of A on A n G is still defined by: a · (b, g) = (aba−1 , a · g). The cocycles associated to a homomorphism ϕ ∈ Hom(Γ, A) are now the maps α : Γ → G that satisfy α(γδ) = (ϕ(δ)−1 · α(γ))α(δ) and the coboundaries are the cocycles of the form α(γ) = (ϕ(γ)−1 · g)g −1 . Finally G acts on Z 1 (ϕ) by (g · α)(γ) = (ϕ(γ)−1 · g)α(γ)g −1 .
Bibliography [AK]
S. Adams and A.S. Kechris, Linear algebraic groups and countable Borel equivalence relations, J. Amer. Math. Soc., 13 (2000), 104–137. [AL] S. Adams and R. Lyons, Amenability, Kazhdan’s property and percolation for trees, groups and equivalence relations, Israel J. Math., 75 (1991), 341–370. [A1] O. Ageev, On the genericity of some nonasymptotic dynamic properties, Russ. Math. Surveys, 58 (1) (2003), 173–174. [A2] O. Ageev, The homogenous spectrum problem in ergodic theory, Invent. Math., 160 (2005), 417–446. [A3] O. Ageev, Spectral rigidity of group actions and Kazhdan groups, preprint, 2006. [AW] C.A. Akemann and M. Walter, Unbounded negative definite functions, Canad. J. Math., 33 (4) (1981), 862–871. [ANO] L. Arnold, Nguyen Dinh Cong, and V.I. Oseledets, The essential range of a nonabelian cocycle is not a cohomology invariant, Israel J. Math., 116 (2000), 71–76. [BK] H. Becker and A.S. Kechris, The descriptive set theory of Polish group actions, Cambridge Univ. Press, 1996. [BdlH] E. B´edos et P. de la Harpe, Moyennabilit´e int´erieure des groupes: d´efinitions et exemples, Enseign. Math., 32 (1986), 139–157. [B] B. Bekka, Amenable representations of locally compact groups, Invent. Math., 100 (1990), 383–401. [BdlHV] B. Bekka, P. de la Harpe, and A. Valette, Kazhdan’s property (T), Cambridge Univ. Press, 2008. [BV] B. Bekka and A. Valette, Kazhdan’s property (T) and amenable representations, Math. Z., 212 (1993), 293–299. [BCR] C. Berg, J.P.R. Christensen and P. Ressel, Harmonic analysis on semigroups, Springer, 1984. [BGo] V. Bergelson and A. Gorodnik, Weakly mixing group actions: A brief survey and an example, Modern dynamical systems and applications, Cambridge Univ. Press, 2004, 3–25. [BR] V. Bergelson and J. Rosenblatt, Mixing actions of groups, Ill. J. Math., 32(1) (1988), 65–80. [Be] S. I Bezuglyi, Groups of automorphisms of a measure space and weak equivalence of cocycles, in: Descriptive Set Theory and Dynamical Systems, London Math. Soc. Lecture Note Series, 277, Cambridge Univ. Press, 2000. [BG1] S.I. Bezuglyi and V. Ya. Golodets, Topological properties of complete groups of automorphisms of a measure space, Siberian Math. J., 21 (1980), 147–155. [BG2] S.I. Bezuglyi and V. Ya. Golodets, Outer conjugacy of actions of countable amenable groups on a space with measure, Math. USSR-Izv., 29 (4) (1987), 1–18. [Bo] L. Bowen, Weak isomorphisms between Bernoulli shifts, arXiv e-print, math.DS/0812.2718v1. [Br] R.M. Bryant, Groups with the minimal condition on stabilizers, J. Algebra, 60 (1979), 371–383. 227
228
BIBLIOGRAPHY
[CCJJV] P.-A. Cherix, M. Cowling, P. Jolissaint, P. Julg, and A. Valette, Groups with the Haagerup property, Birkh¨ auser, 2001. [CN] J. R. Choksi and M.G. Nadkarni, Baire category in spaces of measures, unitary operators and transformations, Inv. subspaces and allied topics, Narosa, 1990, 147–163. [CP] J.R. Choksi and V.S. Prasad, Ergodic theory on homogeneous measure algebras, Measure theory, Oberwolfach 1981, Springer, 1982, 366–408. [CJ1] A. Connes and V.F.R. Jones, A II1 factor with two nonconjugate Cartan subalgebras, Bull. Amer. Math. Soc., 6 (1982), 211–212. [CJ2] A. Connes and V.F.R. Jones, Property (T) for von Neumann algebras, Bull. London Math. Soc., 17(1) (1985), 57–62. [CK] A. Connes and W. Krieger, Measure space automorphisms, the normalizers of their full groups and approximate finiteness, J. Funct. Anal., 24 (1977), 336–352. [CW] A. Connes and B. Weiss, Property T and almost invariant sequences, Israel J. Math., 37 (1980), 209–210. [CFS] I. Cornfeld, S.V. Fomin, and Ya. G. Sinai, Ergodic theory, Springer, 1982. [Da] A. I. Danilenko, The topological structure of Polish groups and groupoids of measure preserving transformations, Publ. RIMS, Kyoto Univ., 31 (1995), 913– 940. [dlH] P. de la Harpe, On simplicity of reduced C ∗ -algebras of groups, Bull. London Math. Soc., 39 (2007), 1–26. [Di] J. Dixmier, C ∗ -Algebras, North Holland, 1982. [DJK] R. Dougherty, S. Jackson and A.S. Kechris, The structure of hyperfinite Borel equivalence relations, Trans. Amer. Math. Soc., 341 (1) (1994), 193–225. [Dy] H.A. Dye, On the ergodic mixing theorem, Trans. Amer. Math. Soc., 118 (1969), 123–130. [Ef1] E.G. Effros, Property Γ and inner amenability, Proc. Amer. Math. Soc., 47(2) 1975, 483–486. [Ef2] E.G. Effros, Polish transformation groups and classification problems, General topology and modern analysis, Acad. Press, 1981, 217–227. [Ei1] S.J. Eigen, On the simplicity of the full group of ergodic transformations, Israel J. Math., 40(3–4) (1981), 345–349. [Ei2] S. J. Eigen, The group of measure-preserving transformations of [0,1] has no outer automorphisms, Math. Ann., 259 (1982), 259–270. [E] I. Epstein, Orbit inequivalent actions of non-amenable groups, arXiv e-print, math.GR/0707.4215v2, 2008. [Fa] A. Fathi, Le groupe des transformations de [0,1] qui preservant la mesure de Lebesgue est un groupe simple, Israel J. Math., 29 (2–3) (1978), 302–308. [FM] J. Feldman and C.C. Moore, Ergodic equivalence relations, cohomology and von Neumann algebras, I, Trans. Amer. Math. Soc., 234(2) (1977), 289–324. [FJ] R.J. Fleming and J.E. Jamison, Isometries on Banach spaces, Chapman and Hall/CRC, 2003. [Fo] G.B. Folland, A Course in Abstract Harmonic Analysis, CRC Press, 1995. [FRW] M. Foreman, D.J. Rudolph and B. Weiss, On the conjugacy relation in ergodic theory, C.R. Math. Acad. Sci. Paris, 343(10) (2006), 653–656. [FW] M. Foreman and B. Weiss, An anti-classification theorem for ergodic measure preserving transformations, J. Eur. Math. Soc., 6 (2004), 277–292. [Fu1] A. Furman, Outer automorphism groups of some ergodic equivalence relations, Comment. Math. Helv., 80 (2005), 157–196. [Fu2] A. Furman, On Popa’s cocycle superrigidity theorem, Int. Math. Res. Not. IMRN, (2007), ID 073.rnm, 46 pages. [Fur] H. Furstenberg, Recurrence in Ergodic Theory and Combinatorial Number Theory, Princeton Univ. Press, 1981.
BIBLIOGRAPHY
[Ga1] [Ga2] [GPo] [Ge]
[GG]
[GP] [Gl1] [Gl2] [GK] [GTW] [GW] [Go] [Gol] [Goo] [GM] [Gr] [Ha] [HO] [HM] [Hj1] [Hj2] [Hj3] [Hj4] [Hj5] [HK1] [HK2]
229
D. Gaboriau, Coˆ ut des relations d’equivalence et des groupes, Invent. Math., 139 (2000), 41–98. D. Gaboriau, Examples of groups that are measure equivalent to the free group, Erg. Theory and Dynam. Syst., 25 (2005), 1809–1827. D. Gaboriau and S. Popa, An uncountable family of nonorbit equivalent actions of Fn , J. Amer. Math. Soc., 18 (2005), 547–559. S.L. Gefter, Outer automorphism group of the ergodic equivalence relation generated by translations of dense subgroup of compact group on its homogeneous space, Publ. RIMS, Kyoto Univ., 32 (1996), 517–538. S.L. Gefter and V. Ya. Golodets, Fundamental groups for ergodic actions and actions with unit fundamental groups, Publ. RIMS, Kyoto Univ., 24 (1988), 821–847. T. Giordano and V. Pestov, Some extremely amenable groups related to operator algebras and ergodic theory, J. Inst. Math. Jussieu, 6(2) (2007), 279–315. E. Glasner, On minimal actions of Polish groups, Top. Appl., 85 (1998), 119–125. E. Glasner, Ergodic theory via joinings, Amer. Math. Society, 2003. E. Glasner and J. King, A zero-one law for dynamical properties, Contemp. Math., 215 (1998), 231–242. E. Glasner, J.-P. Thouvenot and B. Weiss, Every countable group has the weak Rohlin property, Bull. London Math. Soc, 38(6) (2006), 932–936. E. Glasner and B. Weiss, Kazhdan’s property T and the geometry of the collection of invariant measures, Geom. Funct. Anal., 7 (1997), 917–935. R. Godement, Les fonctions de type positif et la theorie des groups, Trans. Amer. Math. Soc., 63 (1948), 1–84. V. Ya. Golodets, Action of T-groups on Lebesgue spaces and properties of full factors of Type II1 , Publ. RIMS, Kyoto Univ., 22 (1986), 613–636. G.R. Goodson, A survey of recent results in the spectral theory of ergodic dynamical systems, J. Dynam. and Control Systems, 5(2) (1999), 173–226. M. Gromov and V.D. Milman, A topological application of the isoperimetric inequality, Amer. J. Math., 105 (1983), 843–854. R. Grzaslewicz, Density theorem for measurable transformations, Colloq. Math., 48 (1984), 245–250. P.R. Halmos, Lectures on Ergodic Theory, Chelsea, 1958. T. Hamachi and M. Osikawa, Ergodic groups of automorphisms and Krieger’s theorems, Sem. Math. Sci., Keio Univ., 3 (1981), 113 pp. S. Hartman and J. Mycielski, On the embedding of topological groups into connected topological groups, Colloq. Math., 5 (1958), 167–169. G. Hjorth, On invariants for measure preserving transformations, Fund. Math., 169 (2001), 1058–1073. G. Hjorth, Classification and orbit equivalence relations, Amer. Math. Society, 2000. G. Hjorth, Non-smooth infinite dimensional group representations, preprint, 1997 (www.math.ucla.edu/∼greg/research.html). G. Hjorth, A converse to Dye’s theorem, Trans. Amer. Math. Soc., 357 (2005), 3083–3103. G. Hjorth, Mixing actions of groups with the Haagerup approximation property, Fund. Math., 203(1) (2009), 47–56. G. Hjorth and A.S. Kechris, Analytic equivalence relations and Ulm-type classifications, J. Symb. Logic, 60 (1995), 1273–1300. G. Hjorth and A.S. Kechris, Borel equivalence relations and classifications of countable models, Ann. Pure Appl. Logic, 82 (1996), 221—272.
230
[HK3]
[HMP] [I1] [I2] [IKT] [IPP]
[JKL] [JP] [Ja] [Jo1] [Jo2] [Jo3] [JS] [KSa] [KTh]
[Kea] [Kec1] [Kec2] [Kec3] [Kec4] [KL] [KLy] [KM] [KR]
[KS] [KT]
BIBLIOGRAPHY
G. Hjorth and A.S. Kechris, Rigidity theorems for actions of product groups and countable Borel equivalence relations, Memoirs Amer. Math. Soc., Vol. 177, No. 833, 2005. B. Host, J.-F. M´ela, F. Parreau, Analyse harmonique des mesures, Asterisque, 135–136, 1988. A. Ioana, Orbit inequivalent actions for groups containing a copy of F2 , arXiv e-print, math.GR/0701027, 2007. A. Ioana, Cocycle superrigidity for profinite actions of property (T) groups, arXiv e-print, math.GR/0805.2998, 2008. A. Ioana, A.S. Kechris and T. Tsankov, Subequivalence relations and positivedefinite functions, Groups, Geometry and Dynamics, to appear. A. Ioana, J. Peterson, and S. Popa, Amalgamated free products of weakly rigid factors and calculation of their symmetry groups, Acta Math., 200(1) (2008), 85–153. S. Jackson, A.S. Kechris, and A. Louveau, Countable Borel equivalence relations, J. Math. Logic, 2(1) (2002), 1–80. J. Jacod and P. Protter, Probability essentials, Springer, 2000. S. Janson, Gaussian Hilbert spaces, Cambridge Univ. Press, 1997. P. Jolissaint, Property (T) for discrete groups in terms of their regular representation, Math. Ann., 297 (1993), 539–551. P. Jolissaint, Borel cocycles, approximation properties and relative property (T), Erg. Theory and Dynam. Syst., 20 (2000), 483–499. P. Jolissaint, On property (T) for pairs of topological groups, Enseign. Math., 51 (2005), 31–45. V.F.R. Jones and K. Schmidt, Asymptotically invariant sequences and approximate finiteness, Amer. J. Math., 109 (1987), 91–114. J.-P. Kahane et R. Salem, Ensembles parfaits et s´eries trigonom´etriques, Hermann, 1994. A. Katok and J.-P. Thouvenot, Spectral properties and combinatorial constructions in ergodic theory, Handbook of Dynamical Systems, Vol. 1B, 649–673, Elsevier, 2005. M. Keane, Contractibility of the automorphism group of a nonatomic measure space, Proc. Amer. Math. Soc., 26 (1970), 420–422. A.S. Kechris, Lectures on definable group actions and equivalence relations, preprint, 1994. A.S. Kechris, Classical descriptive set theory, Springer, 1995. A.S. Kechris, Actions of Polish groups and classification problems, in Analysis and Logic, Cambridge Univ. Press, 2002, 115–187. A.S. Kechris, Unitary representations and modular actions, J. Math. Sciences, 140(3) (2007), 398–425. A.S. Kechris and A. Louveau, Descriptive set theory and the structure of sets of uniqueness, Cambridge Univ. Press, 1989. A.S. Kechris and R. Lyons, Ordinal rankings on measures annihilating thin sets, Trans. Amer. Math. Soc., 310 (1988), 747–758. A.S. Kechris and B.D. Miller, Topics in orbit equivalence, Springer, 2004. A.S. Kechris and C. Rosendal, Turbulence, amalgamation and generic automorphisms of homogeneous structures, Proc. London Math. Soc., 94(3) (2007), 302–350. A.S. Kechris and N.E. Sofronidis, A strong generic ergodicity property of unitary and self-adjoint operators, Erg. Theory and Dynam. Syst., 21 (2001), 1459–1479. A.S. Kechris and T. Tsankov, Amenable actions and almost invariant sets, Proc. Amer. Math. Soc., 136(2) (2008), 687–697.
BIBLIOGRAPHY
[KP] [Ki] [KiT] [Kh] [Le] [LP] [LS] [Lu] [M] [Ma] [Mi] [Mo] [Na] [Nh] [OW] [O] [Pa] [PS] [Pe1] [Pe2] [Po1] [Po2] [Po3] [Po4] [PV] [Poz] [Pr] [Qu]
231
D. Kerr and M. Pichot, Asymptotic abelianess, weak mixing, and property T, J. Reine. Angew. Math., 623 (2008), 213–235. Y. Kida, Orbit equivalence rigidity for ergodic actions of the mapping class group, Geom. Dedicata, 131 (2008), 99–109. J. Kittrell and T. Tsankov, Topological properties of full groups, Erg. Theory and Dynam. Syst., to appear. D. Koshnevisan, Multiparameter processes: An introduction to random fields, Springer, 2002. M. Lema´ nczyk, Introduction to ergodic theory from the point of view of spectral theory, KAIST, 1995. L.-A. Lindahl and F. Poulsen, Thin sets in harmonic analysis, Marcel Dekker, 1971. A Lubotzky and Y. Shalom, Finite representations in the unitary dual and Ramanujan groups, Contemp. Math., 347 (2004), 173–189. W. L¨ uck, The type of the classifying space for a family of subgroups, J. Pure Appl. Algebra, 149 (2000), 177–203. G.W. Mackey, Point realization of transformation groups, Ill. J. Math., 6 (1962), 327–335. B. Maurey, Constructions de suites sym´etriques, C.R. Acad. Sc. Paris, S´er. A–B, 288 (1979), 679–681. B.D. Miller, Full groups, classification and equivalence relations, Ph.D. Thesis, U.C. Berkeley, 2004. C.C. Moore, Group extensions and cohomology for locally compact groups. III, Trans. Amer. Math. Soc., 221(1) (1976), 1–33. M.G. Nadkarni, Spectral theory of dynamical systems, Birkh¨ auser, 1998. N.T. Nhu, The group of measure preserving transformations of the unit interval is an absolute retract, Proc. Amer. Math. Soc., 110 (1990), 515–522. D.S. Ornstein and B. Weiss, Ergodic theory of amenable group actions, I. The Rohlin Lemma, Bull. Amer. Math. Soc. (NS), 2 (1980), 161–164. N. Ozawa, There is no separable universal II1 -factor, Proc. Amer. Math. Soc., 132 (2004), 487– 490. W. Parry, Topics in ergodic theory, Cambridge Univ. Press, 1981. K.R. Parthasarathy and K. Schmidt, On the cohomology of a hyperfinite action, Mh. Math., 84 (1977), 37–48. V. Pestov, Ramsey-Milman phenomenon, Urysohn metric spaces, and extremely amenable groups, Israel J. Math., 127 (2002), 317–357 V. Pestov, Dynamics of Infinite-dimensional Groups, University Lecture Series, Vol. 40, Amer. Math. Soc., 2006. S. Popa, Correspondences, INCREST Preprint, 56/1986. S. Popa, On a class of type II1 factors with Betti numbers invariant, Ann. of Math., 163 (2006), 809–899. S. Popa, Cocycle and orbit equivalence superrigidity for malleable actions of w-rigid groups, Invent. Math., 170(2) (2007), 243–295. S. Popa, Some computations of 1-cohomology groups and constructions on non orbit equivalent actions, J. Inst. Math. Jussieu, 5 (2006), 309–332. S. Popa and S. Vaes, Strong rigidity of generalized Bernoulli actions and computations of their symmetry groups, Adv. Math, 217(2) (2008), 833–872. T. Poznansky, Characterization of linear groups whose reduced C ∗ -algebras are simple, arXiv e-print, math GR/0812.2486v3. V.S. Prasad, Generating dense subgroups of measure preserving transformations, Proc. Amer. Math. Soc., 83 (2) (1981), 286–288. M. Queff´elec, Substitution dynamical systems, Springer, 1987.
232
[R] [Ro] [Ru1] [Ru2] [Rud] [Ry1] [Ry2] [Sc1] [Sc2]
[Sc3]
[Sc4] [Sc5] [SU] [So] [Sun] [Th] [Tho] [To1] [To2] [To3] [V] [Wa] [We] [Y] [Zi1] [Zi2]
BIBLIOGRAPHY
A.G. Robertson, Property (T) for II1 factors and unitary representations of Kazhdan groups, Math. Ann., 296 (1993), 547–555. C. Rosendal, The generic isometry and measure preserving homeomorphism are conjugate to their powers, Fund. Math., to appear. W. Rudin, Fourier-Stieltjes transforms of measures on independent sets, Bull. Amer. Math. Soc., 66 (1960), 199–202. W. Rudin, Fourier analysis on groups, Wiley, 1990. D. J. Rudolph, Residuality and orbit equivalence, Contemp. Math., 215 (1998), 243–254. V.V. Ryzhikov, Representation of transformations preserving the Lebesgue measure in the form of periodic transformations, Math. Notes, 38 (1985), 978–981. V.V. Ryzhikov, Transformations having homogeneous spectra, J. Dyn. Control Syst., 5 (1999), 145–148. K. Schmidt, Cocycles of Ergodic Transformation Groups, MacMillan, 1977. K. Schmidt, Some solved and unsolved problems concerning orbit equivalence of countable group actions, Proc. conf. erg. theory and related topics, II, TeubnerTexte Math., 94, Teubner, 1987, 171–184. K. Schmidt, Amenability, Kazhdan’s property T, strong ergodicity and invariant means for ergodic group actions, Erg. Theory and Dynam. Syst., 1 (1981), 223– 236. K. Schmidt, Asymptotic properties of unitary representations and mixing, Proc. London Math. Soc., 48 (1984), 445–460. K. Schmidt, Algebraic ideas in ergodic theory, CBMS Regional Conference Series in Mathematics, 76, Amer. Math. Soc., 1990. J. Schreier and S. Ulam, Sur le nombre des g´en´erateurs d’un groupe topologique compact et connexe, Fund. Math., 24 (1935), 302–304. S. Solecki, Actions of non-compact and non-locally compact Polish groups, J. Symb. Logic, 65(4) (2000), 1881–1894. V.S. Sunder, An invitation to von Neumann algebras, Springer, 1987. ¨ E. Thoma, Uber unit¨ are Darstellungen abz¨ ahlbarer discreter Gruppen, Math. Ann., 153 (1964), 111–138. S. Thomas, The automorphism tower problem, preprint, 2006. A. T¨ ornquist, Orbit equivalence and actions of Fn , J. Symb. Logic, 71(1) (2006), 265–282. A. T¨ ornquist, Conjugacy, orbit equivalence and classification for measure preserving group actions, Erg. Theory and Dynam. Syst., to appear. A. T¨ ornquist, Localized cohomology and some applications of Popa’s cocycle super-rigidity theorem, arXiv e-print, math.LO/0711.0158v1. J. van Mill, Infinite-dimensional topology, North Holland, 1989. P. Walters, An introduction to ergodic theory, Springer, 1982. T.-J. Wei, Descriptive properties of measure preserving actions and the associated unitary representations, Ph.D. Thesis, Caltech, 2005. H. Yoshizawa, Some remarks on unitary representations of the free group, Osaka Math. J., 3 (1951), 55–63. R. Zimmer, Random walks on compact groups and the existence of cocycles, Israel J. Math., 26(1) (1977), 84–90. R. Zimmer, Ergodic theory and semisimple groups, Birkh¨ auser, 1984.
Index
C(a), 75 C[E] , 25 C(Γ), 54, 75 C ∗ (Γ), 75 C ∗∗ (Γ), 75 CΓ , 47 Cµ (E), 48 CT , 22 C[T ] , 23 CInd∆ Γ (a), 73 CH, 155 D, 11 D(Γ, X, µ), 109 D⊥ , 11 Da , 112 ∆, 1, 99 δu0 , 3 0 , 61 δΓ,u δu , 3 δw , 1 δN [E] , 41 δΓ,∞ , 103 δΓ,u , 61 δΓ,w , 61 δ¯w , 1 ˆ 199 ∆, ˜ 126 d, d˜1 , 148, 149 d˜1∞ , 181 d(xH, yH), 144 dG = d, 125 dµ (A, B), 1 dX,µ,G (f, g), 125 DR, 216 E ∼B F , 35 E∼ = F , 20 E∼ =B F , 29 E ≤B F , 35, 170 E vB F , 29 E ∨ F , 99
A(Γ), 61 A(Γ, X, µ), 61 AZ 1 (Γ, X, µ, G), 130 A n G, 223, 225 [α]∼ , 130, 132 α ∼ β, 130, 132, 224 α ∼w β, 131, 132 α∗ (γ, x), 132 α0 , 163, 175 αT , 177 απ , 165 ∀∞ , 58, 144 α, ˆ 165 a0 , 114 aπ , 198 a∼ =w b, 69 a∼ = b, 61 a ≡ b, 62 a ≺ b, 64 a b, 77 a ∼ b, 64 Aut(G), 224 Aut(X, µ), 1 (Aut(X, µ), w), 1 AMEN(Γ, H), 219 APER, 6 B 1 (E, G), 132 B 1 (ϕ), 224 B 1 (a, G), 130 β A , 134 b ×β (Z, σ), 69 b v a, 69 Bor(X), 199 C-measurable, 123 C(Γ, X, µ), 109 C ∗ -simple, 217 CG (A), 168 Ca , 103, 131 c0 -representation, 79, 217 C⊥ , 2 233
234
EΓX , 13 E0 , 16, 156 E0 -ergodic, 32, 62 E1 , 98 Y EH , 136 ET , 17 Eα , 137 E∞ , 30 Ea , 62 [E], 13 [[E]], 99 ∃∞ , 144 E, 33, 152 ER(α), 141 ER(α), 142 ERG(Γ, H), 217 ERG(Γ, X, µ), 62 ESIM(Γ), 63 E0 RG(Γ, X, µ), 62 ERG, 9 : f1 . . . fn :, 193 Fc (G), 169 Fm , 26 Ft (G), 156 Fc,0 (G), 172 Ft,0 (G), 156 ΦN , 155 ϕh , 199 ϕπ,h , 199 f.a.m., 52 FED(Γ, X, µ), 70 FR(Γ, X, µ), 62 FRERG(Γ, X, µ), 62 Gp , 144 GΓ,H , 95 ˜ 126 G, ˜ α , 151 G G, 142 0 GHAP , 162 Gcount , 161 Γ∗ , 55 γ a , 61 GNS, 195 H 1 (E, G), 132 H 1 (ϕ), 224 H 1 (a, G), 130 H :n: , 193 H ∞ , 190 H n , 189 H ⊗n , 189 HC , 187 HR , 187 ηG , 139
INDEX
ηΓ , 129 ˜ 188 H, H, 219 Hom(Γ, K), 224 HAP, 79 ∞ · ρ, 196 ιT , 128 id, 3 iT , 41 Ind∆ Γ (π), 203 Ind∆ Γ (a), 203 Irr(Γ, H), 95, 210 ICC, 52 κa , 66 κa0 , 66 κ0T , 10 L(X, µ, G), 125 L ∞ , 190 L n , 190 L⊗n , 190 L20 (X, µ), 2 L2 (E, M ), 42 L2 (X, µ), 1 L2 (X, µ, C), 1 L2 (X, µ, R), 1 λΓ , 196, 216 LC, 212 M , 42 MT0 , 92 M0 , 38 µ ˆ(δ), 199 µ ν, 35 µ ∼ ν, 35 µA , 134 µG , 139 µϕ , 191 µh , 199 µπ,h , 199 M, 82 MC , 168 MALGµ , 1 MINCOST(Γ, X, µ), 75 MIX(Γ, H), 217 MIX(Γ, X, µ), 62 MMIX(Γ, X, µ), 62 MMIX(Γ, H), 217 MIX, 10 MMIX, 10 N (0, σ 2 ), 191 N (0, 1), 191 N ER(α), 142 n · π, 208 N [E], 40
INDEX
NLC(Γ, H), 212 1Γ , 68, 79, 196 O(H), 187 OS , 194 O(T, U , V ), 31 Out(E), 44 OE, 62 π∼ = ρ, 95, 207 π ≤ ρ, 95, 208 π ≺ ρ, 208 π ≺Z ρ, 209 π ∼ ρ, 208 πα , 139 πµ , 199 ≺Z , 67 pα , 136, 139 π∆,Γ , 72 P (T), 10 P (Y ), 35 Pr(H), 199 (Q, )-invariant, 87 R, 11 Rα , 135, 139 ρ(A, B), 145 ρΓ,∆ , 74 r, 167 Rep(Γ, H), 95, 207 RepO (Γ, H :1: ), 80 ROUGH(E, G), 150 ROUGH(a, G), 150 ROUGH, 147 S(π), 212 σπ , 200 sΓ , 78, 91 s, 167 Sα , 139 σT0 , 10 σ ˆ (n), 11, 36 σ ∗n , 11 supp(T ), 3 supp(σ), 38 SIM(Γ), 63 SIMν (Γ), 63 SMOOTH(E, G), 150 SMOOTH(a, G), 150 SMOOTH, 147 T × T (x, y), 42 TA , 12 Tσ , 11, 36 TC , 188 [T ], 17 τ˜, 126 , 189
235
⊗, 189, 218 τ (dX,µ,G ), 125 τX,µ,G , 126 T πT −1 , 207 T aT −1 , 61 t(G), 26 τN [E] , 41 u, 3 U (H), 8, 207 U (L2 (X, µ)), 2 UT ×T , 42 UT0 , 2 UT , 2 w, 1 WMIX(Γ, H), 217 WMIX(Γ, X, µ), 62 WMIX, 10 X0 (G), 156 Xα , 135, 139 x 4 y, 65, 145 x ≈ y, 65, 145 χA , 3 (X, µ), 1 Z 1 (E, G), 131 Z 1 (ϕ), 224 Z 1 (a, G), 129 admits non-0 almost invariant vectors, 68 admits non-trivial almost invariant sets, 67 affine automorphism, 223 aperiodic, 5 aperiodic equivalence relation, 13 associated cocycle, 136 automorphic part, 223 Bochner’s Theorem, 199 Borel bireducible, 35 Borel homomorphism, 138 Borel reduced, 35, 171 boson Fock space, 190 centralizer, 103, 168 characteristic function, 3 classified by countable structures, 35 co-induced action, 72 coboundary, 130, 224 cocycle, 69, 129, 131, 224 cocycle identity, 129, 131 cocycle superrigid, 161, 162 cohomologous, 224 cohomologous cocycles, 130, 132 cohomology class, 130, 132
236
cohomology group, 130, 132, 224 cohomology space, 130, 132, 224 complexification, 187 Conjugacy Lemma, 6, 14 conjugacy shift action, 55 conjugate, 133 conjugate representation, 219 contained, 208 Continuum Hypothesis, 155 convergence in measure, 126 cost, 75 countable equivalence relation, 13 counting measure, 129 cyclic dyadic permutation, 5 cyclic vector, 199 densely representable, 216 Dirichlet sets, 11 dyadic permutation, 5 Effros’ Theorem, 144 equivalent cocycles, 130, 132 equivalent measures, 35 ergodic, 62, 217 ergodic equivalence relation, 13 essential range, 141 factor, 69 finite equivalence relation, 13 finite factors, 175 first chaos, 193 fixed price, 75 free action, 25 full group, 13 Gaussian centered measure, 191 Gaussian Hilbert spaces, 193 Gaussian shift, 192 Haagerup Approximation Property, 79 homomorphism, 68 hyperfinite equivalence relation, 18 independent set, 37 induced action, 74, 203 induced representation, 203 induced transformation, 12 infinite conjugacy classes, 52 inner amenable, 52 inner automorphism group, 13 intertwining operator, 210 irreducible, 210 isomorphic, 20, 61, 207 isomorphism, 95
INDEX
isomorphism cocycle, 140 join, 99 joining, 65 Kazhdan pair, 85 kernel, 137 Koopman representation, 2, 66 locally closed, 212 locally topologically finitely generated, 27 Mackey action, 137, 139 Mackey range, 137, 139 malleable, 184 marginal, 65 maximal spectral type, 200 measure algebra, 1 measure preserving equivalence relation, 13 measure-hyperfinite, 176 mild mixing, 10, 62, 217 minimal condition on centralizers, 168 minimally almost periodic, 83 mixing, 10, 62, 217 mixing representation, 79 multiplicative operators, 2 Mycielski, Kuratowski Theorem, 18 negative-definite, 82 non-0 almost invariant vectors, 79 non-trivial asymptotically invariant sequence, 57 orbit equivalence cocycle, 140 orbit equivalent, 57, 62 orthogonal group, 187 outer automorphism group, 44 periodic, 5 periodic equivalence relation, 13 Poincar´e flow, 137, 139 positive-definite function, 191, 192, 199 positivity preserving operators, 2 product action, 65 projection valued measure, 199 property (FD), 219 property (M), 83 property (S), 83 property (T), 81 property (W), 83 Rajchman measure, 36 realification, 187
INDEX
realizable by an action, 220 realized, 208 recurrent, 140, 145 reduction, 136 reduction cocycle, 138, 139 relative property (T), 115 residually amenable, 27 residually finite, 27 restricted multiplicity, 38 Rokhlin property, 93 Schur’s Lemma, 210 semidirect product, 223 shift action, 46 skew product action, 69 smooth, 28, 138, 139 somewhere cocycle superrigid, 162 spectral measure, 200 spectral rigidity, 93 spectrally equivalent, 7, 66 stabilizer, 131 stabilizer of the action, 47 stable, 56 standard measure space, 1 strict cocycle, 129, 131 strong operator topology, 187 strong topology, 207 strongly ergodic, 32, 62 strongly treeable, 55, 186 subrepresentation, 95 symmetric difference, 1 symmetric Fock space, 190 symmetric tensor power, 189 topological ergodic decomposition, 65 topological generator, 26 topologically locally finite, 6 transient, 140 translation part, 223 treeable, 52 trivial cocycle, 130 trivial one-dimensional representation, 79 turbulent, 30, 45 Uniform Approximation Theorem, 6, 14 uniform topology, 3 unitarily equivalent, 7, 66 unitary equivalence, 95 variety, 107 Weak Approximation Theorem, 5 weak mixing, 10, 62, 217
237
weak operator topology, 187 weak topology, 1, 207 weakly commutative groups, 53 weakly contained, 64, 208 weakly contained in the sense of Zimmer, 209 weakly equivalent, 64, 131, 208 weakly isomorphic, 69 Wick product, 193 Wiener chaos decomposition, 193 witness the weak commutativity, 56